research paper on short term memory

• A-Z Publications

Annual Review of Psychology

Volume 59, 2008, review article, the mind and brain of short-term memory.

• John Jonides 1 , Richard L. Lewis 1 , Derek Evan Nee 1 , Cindy A. Lustig 1 , Marc G. Berman 1 , and Katherine Sledge Moore 1
• View Affiliations Hide Affiliations Affiliations: Department of Psychology, University of Michigan, Ann Arbor, Michigan 48109; email: [email protected]
• Vol. 59:193-224 (Volume publication date January 2008) https://doi.org/10.1146/annurev.psych.59.103006.093615
• © Annual Reviews

The past 10 years have brought near-revolutionary changes in psychological theories about short-term memory, with similarly great advances in the neurosciences. Here, we critically examine the major psychological theories (the “mind”) of short-term memory and how they relate to evidence about underlying brain mechanisms. We focus on three features that must be addressed by any satisfactory theory of short-term memory. First, we examine the evidence for the architecture of short-term memory, with special attention to questions of capacity and how—or whether—short-term memory can be separated from long-term memory. Second, we ask how the components of that architecture enact processes of encoding, maintenance, and retrieval. Third, we describe the debate over the reason about forgetting from short-term memory, whether interference or decay is the cause. We close with a conceptual model tracing the representation of a single item through a short-term memory task, describing the biological mechanisms that might support psychological processes on a moment-by-moment basis as an item is encoded, maintained over a delay with some forgetting, and ultimately retrieved.

Article metrics loading...

Full text loading...

Data & Media loading...

• Article Type: Review Article

Most Read This Month

Most cited most cited rss feed, job burnout, executive functions, social cognitive theory: an agentic perspective, on happiness and human potentials: a review of research on hedonic and eudaimonic well-being, sources of method bias in social science research and recommendations on how to control it, mediation analysis, missing data analysis: making it work in the real world, grounded cognition, personality structure: emergence of the five-factor model, motivational beliefs, values, and goals.

Publication Date: 10 Jan 2008

Online Option

Sign in to access your institutional or personal subscription or get immediate access to your online copy - available in PDF and ePub formats

Advertisement

Can short-term memory be trained?

• Open access
• Published: 27 February 2019
• Volume 47 , pages 1012–1023, ( 2019 )

Cite this article

You have full access to this open access article

• Dennis G. Norris 1 ,
• Jane Hall 1 &
• Susan E. Gathercole 1  

15k Accesses

19 Citations

1 Altmetric

Explore all metrics

Is the capacity of short-term memory fixed, or does it improve with practice? It is already known that training on complex working memory tasks is more likely to transfer to untrained tasks with similar properties, but this approach has not been extended to the more basic short-term memory system responsible for verbal serial recall. Here we investigated this with adaptive training algorithms widely applied in working memory training. Serial recall of visually presented digits was found to improve over the course of 20 training sessions, but this improvement did not extend to recall of either spoken digits or visually presented letters. In contrast, training on a nonserial visual short-term memory color change detection task did transfer to a line orientation change detection task. We suggest that training only generates substantial transfer when the unfamiliar demands of the training activities require the development of novel routines that can then be applied to untrained versions of the same paradigm (Gathercole, Dunning, Holmes, & Norris, 2019 ). In contrast, serial recall of digits is fully supported by the existing verbal short-term memory system and does not require the development of new routines.

Similar content being viewed by others

Is long-term memory used in a visuo-spatial change-detection paradigm?

Control processes in short-term storage: Retrieval strategies in immediate recall depend upon the number of words to be recalled

Training change detection leads to substantial task-specific improvement.

Avoid common mistakes on your manuscript.

In this study, we asked whether the capacity of short-term memory (STM) can be improved with practice. In recent years there has been strong interest in the potential of cognitive training programs to enhance mental capacity (Bavelier, Green, Pouget, & Schrater, 2012 ; Simons et al., 2016 ) Many training programs employ complex working memory (WM) activities that combine serial recall of memory sequences with other processing demands. For example, participants may be required to engage in distractor activities interpolated between the presentation of memory items (Chein & Morrison, 2010 ) or to continuously update the sequence of memory items to be remembered (Dahlin, Neely, Larsson, Backman, & Nyberg, 2008 ; Jaeggi, Buschkuehl, Jonides, & Perrig, 2008 ). In these programs, the difficulty of the training task adapts as performance improves with practice. After more than a decade of research in this field, the consensus is that this kind of training generates reliable near transfer to untrained WM tasks with similar task demands. However, there is little far transfer to different activities that are also associated with WM, such as attentional control, reasoning, and learning (Cortese et al., 2015 ; Melby-Lervåg & Hulme, 2013 , 2016 ; Simons et al., 2016 ).

To explain this restricted pattern of transfer we have proposed that training on complex WM tasks involves acquiring a new cognitive skill (Gathercole, Dunning, Holmes, & Norris, 2019 ). The suggestion is that to accomplish the unfamiliar tasks present in most WM training programs, trainees must develop novel routines that coordinate the cognitive processes required. Learning a new routine follows the usual course of the acquisition of any cognitive skill (Anderson, 1982 ; Taatgen, 2013 ). With practice, the routine becomes more efficient and less demanding of general cognitive resources, and performance improves. Transfer will only occur if the routine can be successfully applied to a new task, and this will only happen if the task demands are closely matched. Consistent with this framework, a meta-analysis of WM training studies showed that substantial transfer following WM training is largely restricted to cases in which both the trained and untrained tasks share the same complex WM paradigm (Gathercole et al., 2019 ).

We have made the strong claim that new routines will only be developed if existing mechanisms and processes are not available to support the training activities. If existing mechanisms are available, a new routine is not required, and there will therefore be little transfer. It is proposed that verbal STM measures such as digit span are examples of tasks that do not demand a new routine (Gathercole et al., 2019 ). Highly specialized processes in verbal STM are responsible for the encoding, maintenance, and retrieval of phonological material in its original sequence, with key phenomena being successfully simulated by computational models incorporating separate item- and order-encoding mechanisms (Hurlstone, Hitch, & Baddeley, 2014 ). These processes are frequently engaged in everyday situations outside of the laboratory, including learning new words (Baddeley, Gathercole, & Papagno, 1998 ; Gathercole, 2006 ), following complex verbal instructions (Engle, Carullo, & Collins, 1991 ; Jaroslawska, Gathercole, Logie, & Holmes, 2016 ), and performing mental arithmetic (Adams & Hitch, 1997 ; Geary, Hoard, Byrd-Craven, & DeSoto, 2004 ; McLean & Hitch, 1999 ). On this basis, the cognitive-routine framework predicts that, in contrast to complex WM paradigms, training on verbal STM measures will generate little transfer to other verbal serial-recall tasks.

Surprisingly little research has applied the adaptive computerized algorithms widely employed in WM training studies to more simple verbal STM tasks. Older studies involving small numbers of individuals training on single tasks over extended periods have demonstrated that performance on digit serial recall can improve with practice. In the first study of its kind, two adults received more than 50 sessions of digit span testing spread across a period of four months (Martin & Fernberger, 1929 ). Span increased by about 40% in both cases, and this was accompanied by reports of grouping strategies. Striking evidence that memory span gains across training are driven by mnemonic strategies was provided by a study by Chase and Ericsson ( 1981 ) of S.F., an adult whose digit span increased from seven to 79 items over 2 years of practice. This was achieved by recoding digit sequences into long-distance running times that he was familiar with as a long-distance runner. However, the increase in S.F.’s memory span was entirely restricted to digit sequences. When the memory sequences were composed of letters, his span remained at seven across the full training period.

A second long-distance runner, D.D., was instructed to use the same strategy S.F. had developed, and his span also increased substantially (for a detailed analysis of D.D.’s recall strategies, see Ericsson & Staszewski, 1989 ; Staszewski, 1990 ; Yoon, Ericsson, & Donatelli, 2018 ). Other studies have also demonstrated that digit span expands when participants use other elaborate encoding and retrieval strategies, such as the method of loci, the method of mental imagery, and associations between digit sequences and famous historical dates (Kliegl, Smith, Heckhausen, & Baltes, 1987 , and Susukita, 1933, cited in Kliegl et al., 1987 ). Adopting a rather different approach, Reisberg, Rappaport, and O’Shaughnessy ( 1984 ) trained participants to map digits onto finger movements. In this study, digit span increased by up to 50%.

In each of these cases, the increase in span was accompanied by the use of complex mnemonic strategies that combine existing knowledge with either sequences of multiple digits or well-learned sequences that could be used as cues for retrieval. The gains therefore appear to be a consequence of using long-term memory to support the encoding and retrieval of memory items, in conjunction with fixed-capacity verbal STM (Norris, 2017 ). There is little evidence for fundamental capacity changes in verbal STM with extensive practice.

In the only study of its kind, to our knowledge, Harrison et al. ( 2013 ) employed an adaptive computerized STM training regime that consisted of two simple serial tasks—letter span and a spatial span task involving recall of the spatial locations of cells highlighted successively in a matrix. They compared simple span training with two adaptive training programs employing complex span and visual search tasks. Although performance on untrained STM tasks of word span (a verbal serial-recall task) and arrow span (spatial serial recall) improved following training, equivalent benefits were also found for visual search training. There was therefore no selective enhancement of simple memory span by serial-recall training. Other studies have also failed to detect significant transfer of the Cogmed WM training program to digit span, despite its inclusion of a letter span task included in a small number of training sessions (Brehmer, Westerberg, & Bäckman, 2012 ; Dunning & Holmes, 2014 ; Gray et al., 2012 ; Hardy, Willard, Allen, & Bonner, 2013 ).

In the present study, we tested directly whether the capacity of verbal STM can be enhanced through task-specific training of verbal serial recall. Transfer to untrained STM tasks was compared in three groups, each receiving adaptive training on one of the following STM tasks: digit span, circle span, and color change detection. Circle span involves the serial recall of spatial locations highlighted in a sequence at presentation (Minear et al., 2016 ). Like the dot matrix (Alloway, Gathercole, Kirkwood, & Elliott, 2008 ), Corsi block (Darling, Della Sala, Logie, & Cantagallo, 2006 ), and span-board (Wechsler, 1981 ) tasks, circle span is considered to tap a limited-capacity visuospatial STM system (Logie & Pearson, 1997 ). The color change detection task has been widely used as a measure of the capacity of visual STM (e.g., Awh, Barton, & Vogel, 2007 ). Developed by Luck and Vogel ( 1997 ), it involves participants detecting changes in the colors of individual squares presented simultaneously and briefly in a multi-item visual display. It provides the ideal active-control training condition for the two serial-recall training conditions of digit and circle span, as it does not require the retention of serial order. Performance on this task has already been shown to improve with training. With an adaptive algorithm, Buschkuehl, Jaeggi, Mueller, Shah, and Jonides ( 2017 ) reported that set size increased from 6.3 to 8.8 over ten sessions of 300 trials each. In a related nonadaptive task in which only the target square was presented at test, rather than the entire display, Xu, Adam, Fang, and Vogel ( 2018 ) reported an increase in memory capacity from 2.1 to 3.0 across 60 days of training. Harrison et al. ( 2013 ) reported no changes in change detection following training on two complex span tasks involving serial recall. This suggests that the paradigms tap independent STM and WM systems.

The primary aim of the present study was to investigate what changes occur following digit span training. We set out to track transfer by systematically varying the individual features of the trained tasks in a set of untrained tasks administered before and after training. The untrained verbal serial-recall tasks were spoken digit span (a change in presentation modality from the visual trained task) and visual letter span (a change in verbal category). Substantial experimental evidence has shown that each of these input forms gains ready access to the phonological storage component of verbal STM (Baddeley, Lewis, & Vallar, 1984 ).

The design of this study allowed us to address several hypotheses regarding the transferability of STM training. In each case, we did this by comparing the training program of interest with the most appropriate active-control program to test each hypothesis. The first hypothesis was that there would be no transfer following digit span training to other verbal serial-recall tasks, because such tasks depend on specialized processes that are already in place in verbal STM (Gathercole et al., 2019 ). As a consequence, they would not require the development of the novel routines proposed to be the primary source of transfer to untrained tasks with compatible structures. There should therefore be no transfer across verbal serial-recall tasks, even though they place comparable demands on the retention of phonological serial-order information. This should be evident when comparing the impact of digit span training with that of circle span training, the most similar active-control condition, which differed in the domain of STM, but not in the requirement for serial recall. Such an outcome would indicate not only that the fundamental capacity of verbal STM is impervious to training, but also that any substantial on-task training gains are mediated by processes that operate largely outside of STM, such as recoding or chunking (Ericsson, Chase, & Faloon, 1980 ; Martin & Fernberger, 1929 ). Later in this article, we speculate that more subtle improvements may nonetheless result from optimization or fine-tuning of the task model to its unique combination of features.

An alternative hypothesis is that verbal STM can be trained. This would be consistent with claims that the cognitive and neural processes that underpin the broader WM system in which STM is embedded can be modified by intensive training (Astle, Barnes, Baker, Colclough, & Woolrich, 2015 ; Klingberg, 2010 ). If transfer does extend to untrained serial-recall tasks, the key question is: to which tasks? If the serial-order mechanism is both trainable and specific to verbal STM, transfer should not extend beyond verbal serial-recall tasks to any other untrained tasks, including the circle span measure of spatial STM. In this way, the data have the potential to inform long-standing debate about the extent to which the STM mechanisms for retaining serial order are domain-specific or domain-general (Abrahamse, Van Dijck, Majerus, & Fias, 2014 ; Alloway, Gathercole, & Pickering, 2006 ; Bayliss, Jarrold, Gunn, & Baddeley, 2003 ; Engle et al., 1991 ; Hanley, Young, & Pearson, 1991 ; Hurlstone et al., 2014 ; Majerus et al., 2010 ). We asked whether transfer is restricted to serial-recall tasks in the same domain (from digit span to letter span and spoken digit span) or extends across domains (from digit span to circle span or circle span to digit span). Here, the critical comparison to evaluate the specificity of transfer would be digit span training versus nonserial color change training.

To further test the limits on transfer, other untrained tasks were included that did not require memory for serial order. Pattern span involves recall of the pattern of filled cells in a static grid. This measure of visual STM involves the recall of filled cells in a grid that are displayed simultaneously and can be recalled in any order. Under some conditions, performance on this test has been found to be dissociable from serial spatial STM tests such as Corsi block recall, possibly reflecting fractionation of visuospatial STM into separate visual and spatial components (Della Sala, Gray, Baddeley, Allamano, & Wilson, 1999 ). We might therefore anticipate no transfer from either digit span or circle span training (relative to color change detection training) to pattern span.

An untrained line orientation change detection test was also included, in which the array was composed of multiple lines at different orientations and in which participants judged whether the orientation of a single element had changed or remained the same. Alongside color change detection training, this allowed us to test whether training in verbal or spatial serial recall generates benefits that extend to nonserial visual STM. We predicted that it would not, as serial recall appears to reflect a distinct and purely visual system of temporary storage. The inclusion of this training condition also provided an opportunity to explore whether training in one visual change detection task (color change detection) transfers to the ability to detect changes in other visual features in a similar task environment. To date, there has been little research on transfer following training on this paradigm; only a single study has been performed, and this showed no transfer across to variants of the same paradigm with minor changes (Gaspar, Neider, Simons, McCarley, & Kramer, 2013 ). A finding of positive transfer to orientation change detection following color change detection training would provide preliminary evidence for training-related improvement in the ability to detect mismatches in the properties of visual displays, rather than more specifically to detect changes in the colors of individual objects. In the absence of strong hypotheses regarding the outcome, comparisons were made between the color change training group and each of the other three training groups, including a passive control group that received no training, as well as the digit and circle span training groups, which formed the two active controls.

Participants

Eighty English-speaking adults between 18 and 35 years of age were recruited from the Medical Research Council Cognition and Brain Sciences Unit (CBU) Volunteer Panel. Participants received a payment for their time and travel expenses. The study was approved by the University of Cambridge Psychology Research Ethics Committee (CPREC 2014.76). The Participant Information Sheet provided in advance of recruitment outlined a standard hourly payment for participation in the study and reimbursement for travel expenses. It explained that participants would be randomly allocated to conditions, with 10 h of paid home-based iPad training for some but not all individuals. The participants were allocated to training conditions (digit span, circle span, change detection, and no training) on a random basis, subject to the constraint that there were 20 participants in each training condition. This sample size yields power of .91 to detect a large effect size, f 2 = .35 with a general linear regression model (GLM), and .55 to detect a medium effect size, f 2 = .35. The sample size was determined on the basis of the outcomes of a meta-analysis of near transfer following WM training reported by Gathercole et al. ( 2019 ). On the basis of the limited available evidence, they concluded that, in contrast to the robust transfer found following training on complex WM tasks, gains in verbal STM are at best “small in magnitude and may be reliably detected only under conditions of higher statistical power than the standard WM training study” (p. 20). The present sample size of 20 participants per training group falls within the standard range for WM training studies (Dunning & Holmes, 2014 ; Harrison et al., 2013 ; Henry, Messer, & Nash, 2014 ).

Individuals made two 3-h visits to the CBU. On the first visit, each person was given an iPad with a retina display, for use only in the experiment, that provided access to the training program. Participants first completed eight iPad transfer tasks and then eight tests from the Automated Working Memory Assessment (AWMA). The data from the AWMA are reported for completeness only. The training program was then demonstrated, and participants took the iPad away with them to perform the training. The second visit took place shortly after the final session, approximately three weeks later. Participants again completed the iPad and AWMA transfer tests, and then they returned the iPad. All phases of the experiment apart from administration of the AWMA were presented on iPads with a display resolution of 2,048 × 1,536 pixels in landscape mode. At both the pre- and posttraining sessions, there was also a resting-state magnetoencephalography session, the data from which will be reported elsewhere.

Each of the following tasks was administered at the two visits to the CBU, before and after completing the training program. The eight tasks were presented on the iPad, with common designs and structures where possible. After each session of transfer and training, the data were automatically uploaded to a server at the CBU. Two participants did not complete the pretraining line orientation change detection task, and a further participant did not complete the posttraining line orientation task.

Visual digit span

Digits were presented at a rate of one per 750 ms, with each digit being displayed for 500 ms and a blank interval of 250 ms between digits. At the end of the digit sequence, a numeric keyboard (the digits 1–9 in 3 × 3 telephone keypad layout) was displayed, and participants pressed the keys in the order in which the digits had appeared. Below the keyboard was a “Done” key that participants pressed after recall had been completed. With list lengths of nine or less, the digits were sampled randomly without replacement from the digits 1–9. With list lengths greater than nine, the initial set of nine digits was supplemented with a further, randomly sampled N digits. No digits appeared twice in succession, and there were no runs of three or more consecutive ascending or descending digits. Testing began with a block of six trials with a list length of four, and increased by one when participants got four or more of the six trials at that length completely correct. Testing continued until participants failed to reach this continuation threshold. Span was determined to be the longest list length for which four or more lists were recalled correctly. Once span had been determined, 12 further trials were presented at each of the lengths span + 1 and span + 2. The measure of performance was 2 * the number of items recalled correctly during span setting + the numbers of items recalled correctly at span + 1 and span + 2. This gave equal weight to trials at each list length.

Spoken digit span

This task was identical to the visual digit task, except that the stimuli were digits spoken by a male speaker. The digit sound files were padded out with silence to be 500 ms long. Each digit sound file was followed by 250 ms of silence. The method for determining span and the numbers of trials at each of span + 1 and span + 2 were identical to the methods employed for visual digit span.

Letter span

This task was also identical to the visual digit span task, except that the stimulus set was composed of the consonants B, F, H, J, L, M, Q, R, and S. The letters on a 3 × 3 keyboard were arranged in alphabetical order. The method for determining span and the numbers of trials at each of span + 1 and span + 2 were identical to the methods employed for visual digit span.

Circle span

Participants were presented with an array of pseudorandomly positioned circles with a radius of 81 pixels and a minimum center-to-center separation of 272 pixels. All circles were colored medium blue on a gray background and then, in random sequence, each circle turned light blue for 250 ms. The rate of presentation was 750 ms per circle. At the end of the sequence, all circles remained visible, and participants were instructed to touch the circles in the order in which they had been displayed. The method for determining span and the numbers of trials at each of span + 1 and span + 2 were identical to the methods employed for visual digit span. Although this task has been used elsewhere as a test of spatial STM (Minear et al., 2016 ), the possibility cannot be ruled out that participants might attempt to encode the location of the circles verbally. This could introduce an element of verbal STM training in the circle span task. This possibility should be minimized here by the fact that the configuration of the circles varied randomly from trial to trial, making it difficult to assign a set of consistent verbal labels to the locations of the circles.

Pattern span

On each trial a 6 × 6 grid was presented for 500 ms. Initially, four of the cells in the grid were displayed in red. After a delay of 1,000 ms, an empty grid was presented, and participants had to touch the squares that had been presented in red. The squares could be touched in any order. The method for determining span and the numbers of trials at each of span + 1 and span + 2 were identical to the methods employed for visual digit span.

Color change detection

The procedure was based on that used by Luck and Vogel ( 1997 ). Participants saw an array containing between three and 23 colored squares presented for 250 ms. The colors of the squares were chosen at random with replacement from a set of seven readily discriminable colors. The locations of the squares (38 pixels) were random, subject to the constraint that a minimum distance of 117 pixels should separate the centers of the squares.

After a blank retention interval of 1,000 ms, a probe display appeared for 500 ms. The probe was constructed by repeating the previous array, but with one square chosen at random to be the probe square. The color of the probe square either remained the same as in the initial display or, in 50% of the trials, was changed to another randomly chosen color. The location of the probe square was indicated by a red rectangle. Participants had a maximum of 5,000 ms to judge whether the color of the probe square had changed. In all, 20 trials were presented for each of the array sizes 6, 9, and 12. Cowan’s K was used to provide a measure of STM capacity for this task (Cowan, 2001 ; Cowan et al., 2005 ), where K = display size × (proportion hits – proportion false alarms). The mean K was computed over the three array sizes, and this measure was used for the purposes of analysis.

Direction change detection

The orientation change detection task was identical to the color change detection task, except that the colored squares were replaced by black lines that could appear in one of four orientations (vertical, horizontal, or either diagonal). On each probe trial, one line was cued by a red circle, with a 50% probability that the orientation of the line would have changed. The mean K was computed, as in the color change detection task. Due to technical problems, there were incomplete data on this transfer task for two participants in the circle-training condition. Their data were omitted from the reported analyses.

Automated Working Memory Assessment

The following tests from the AWMA (Alloway, 2007 ), a standardized test battery of STM and WM tests, were administered on a desktop PC. Each employed a span procedure. The tests were word span and nonword span tests (verbal STM), dot matrix and mazes memory (visuospatial STM), listening span and counting span (verbal WM), and Mr X and spatial span (visuospatial WM). The analysis was based on raw scores. It should be noted that these tasks shared fewer presentational and task features in common with the most closely matched training activities than did the iPad transfer tasks. On this basis, no strong predictions could be made regarding the impact and potential specificity of the training conditions on these transfer tests. For completeness, the data and statistical outcomes are reported in the supplemental material .

Matrix reasoning

This test of nonverbal reasoning from the Wechsler Adult Intelligence Scales involves selecting the missing part to complete visuospatial patterns. Raw scores were used for the purpose of analysis.

Each participant completed either digit span, circle span, color change detection training, or no training. The participants in the three active training conditions were asked to complete 20 sessions in total on their iPad program. The maximum time allowed for completion of a session was 40 min, with no more than three sessions per day and an interval of no more than two days between successive training sessions. Training could only be performed between 7 a.m. and 11 p.m.

Digit span training

Each training session consisted of eight blocks of ten trials employing the same procedure as the visual digit span transfer test, with the exception that set size was varied adaptively. Training began with a sequence of three digits, and increased by one when participants got eight or more trials completely correct in a block. The length of the sequence decreased by one if participants got two or fewer trials correct. Due to technical problems, the data for one participant in the digit training condition were lost for the 12th session. The missing score was replaced by the mean score from the 11th and 13th sessions. On average, participants completed the training sessions in in 13.8 days (min 10, max 16, SD 1.8). The principal score for the purposes of analysis was the span reached in the final block of each session.

Circle span training

Each training session consisted of eight blocks of ten trials employing the same procedure as the circle span transfer task. Training began with a display of three circles and increased by one when participants got eight or more trials completely correct in a block. The number of circles decreased by one if participants got two or fewer trials correct. On average, participants completed the training sessions in 13.25 days (min 10, max 15, SD 1.55). The principal score for the purposes of analysis was the span reached in the final block of each session.

Color change detection training

In each training session there were eight blocks of 30 trials, employing the same presentation procedure as in the color change detection transfer task. The size of the array was increased by one if participants got 27 or more trials correct, and decreased by one if they got 18 or fewer correct. On average, participants completed the training sessions in 13 days (min 10, max 15, SD 1). Two measures derived from each session were used for the purposes of analysis: The first was the capacity measure K (Cowan, 2001 ; Cowan et al., 2005 ), and the second was the difficulty level (set size) reached by the end of each block.

Analysis plan

The training and transfer data were analyzed using both traditional null-hypothesis significance testing (NHST) methods and corresponding Bayesian methods. This allowed us to quantify the strength of evidence both in favor of the null hypotheses of the absence of training/transfer effects, and in favor of the alternative hypothesis that there were positive effects. The Bayesian analyses were conducted using JASP (JASP Team, 2015 ). Bayes factors (BF 10 ) were interpreted as follows (Jeffreys, 1961 ): BFs < 0.33 provide evidence for the null hypothesis; BFs 0.33–3 provide equivocal evidence for both hypotheses; BFs > 3 provide evidence favoring the alternative hypothesis; BFs > 10 and < 0.01 are considered strong evidence in either direction; and BFs > 100 and < 0.001 provide decisive evidence in either direction.

Training effects for the three active training conditions were analyzed in one-way analyses of variance (ANOVAs) with session as the independent variable. Interactions between training conditions and trials were not computed, due to the different performance metrics used for the span and color change detection tasks. For all training conditions, the metric was the difficulty level achieved by each participant in the final block of each session. For digit and circle span, this was the number of items in the sequence, and for color change detection, it was K . Both Bayesian and non-Bayesian ANOVAs were performed.

To evaluate the specificity of transfer following training, Bayesian and non-Bayesian linear regression analyses were performed for each combination of the training task of interest and each transfer test. The posttraining transfer measure was the dependent variable, and the pretraining measure and the particular training group contrast were entered as dependent variables in each case. Four group contrasts were made for each transfer measure. Three comparisons contrasted pairs of active adaptive training conditions: digit span versus circle span (testing the domain specificity of serial-recall training), digit span versus color change detection (testing the specificity of serial-recall training), and circle span versus color change detection (testing the specificity of change detection training). A final contrast compared color change detection training with no training, as a test of whether there was transfer across serial and nonserial STM paradigms. For the NHST, a Bonferroni correction was applied on a family-wise basis for the transfer tests, yielding an α of .007. For each group contrast and transfer measure combination, initial linear regression analyses were run testing for interactions between pretraining scores and group. Where these were considered to be significant or to favor the alternative hypothesis ( p < .05 or BF > 3), the group term reported here is taken from the analysis that included the interaction term. If the interaction terms did not meet these criteria, the model was rerun excluding the interaction term, and the group term from this analysis is reported.

Training data

The mean scores achieved at the end of each training session (span for digit and circle training, and both capacity K and difficulty level for color change detection) are shown in Fig. 1 . Gains across training sessions were considerably greater for color change detection than for either digit span or circle span training. Performance increased from the first to the final training session, by 18% for digit span training and by 13% for circle span training. For color change detection training, the increase was 51% for the capacity measure K , and 83% for the difficulty level. This reflects an increase in the number of elements in the array from 8.45 to 15.50.

Scores on the final block of each session as a function of training condition and measure

Bayesian and non-Bayesian one-way ANOVAs were performed on the scores in each session for each training condition. In each case, performance increased significantly across training. For NHST, these results were: digit span, F (19, 361) = 7.691, MSE = .497, p < .001; circle span, F (19, 361) = 6.275, MSE = .358, p < .001; color change detection K , F (19, 361) = 8.774, MSE = 1.227, p < .001; color change detection difficulty, F (19, 361) = 50.574, MSE = 1.447, p < .001. Tested against the null model, the BF 10 values were > 100 for digit span, circle span, color change detection K, and color change detection difficulty level. These outcomes provide decisive evidence that performance improved with training on each task and for each measure.

Transfer data

Descriptive statistics and analyses of the transfer measures are shown in Table 1 . Our first question was whether there is domain-specific transfer to other verbal span tasks following digit span training. This was addressed by comparing posttraining scores on the untrained verbal span tests for the digit and circle span training groups. Unsurprisingly, digit span led to a substantially greater enhancement of digit span performance than did circle span training, according to both NHST and Bayesian analysis. This represents a further demonstration of on-task training. For the two untrained verbal span measures, the evidence for a selective advantage following digit span training was weak. For spoken digit span (a change in the modality of the memory items), the NHST was nonsignificant, and the Bayesian outcome was equivocal, weakly favoring the null hypothesis. For letter span, too, the p value was nonsignificant and the Bayes factor value equivocal.

The second question was whether there are domain-general benefits to serial-recall training. This was addressed in two ways. The first was by comparing the digit span and color change detection training groups on the circle span transfer measure. There was no substantial evidence of transfer: The p value was nonsignificant, and the Bayes factor was equivocal, weakly favoring the null hypothesis. The second comparison was between circle span and color change detection training for the three verbal span measures. In each case, the NHST was nonsignificant and the Bayes factor value substantially favored the null hypothesis. We therefore found no substantial evidence for cross-domain transfer across serial-recall tasks in either direction.

The third question was whether there were paradigm-general benefits to training that extended across all three STM training and transfer tasks. This was addressed by comparing the posttraining performance of the color change detection and no-training groups on the five transfer tests that did not involve change detection. It should be noted that this comparison between an active-training and a no-contact control condition was likely to overestimate any potential benefits, and therefore increase the likelihood of a false-positive result (Simons et al., 2016 ). For two of the three verbal span measures (spoken digit span and letter span), the analyses provided no evidence of training benefits, with nonsignificant p values and Bayes factor values showing substantial support for the null hypothesis. For visual digit span, the p value was nonsignificant, and the Bayes factor was equivocal and mildly favored the null hypothesis. For circle span, though, we did find evidence of strong transfer, by both NHST and Bayesian analysis. This provided unexpected evidence for transfer from visual STM training to visuospatial serial recall. However, it is notable that there was no evidence for a symmetrical pattern of transfer from circle span training to the color change detection task: With circle training, capacity K increased from 5.28 to 5.73. For the no-training condition, a similar increase from 5.26 to 5.71 was observed.

The final question was whether color change detection training generates benefits for a line detection task employing the same paradigm. Here the statistical outcomes were clear. Relative to circle span training, color change detection training was associated with greater improvements in the untrained line orientation change detection task, as indicated by a strongly significant p value and a Bayes factor substantially favoring the alternative hypothesis.

Performance improved on all three STM tasks across the course of training. The performance gains across training were relatively small for the two serial-recall tasks—15% for digit span and 12% for circle span training. For color change detection training, the increases were considerably greater, with an estimated STM capacity increase of 51%, and an 83% increase in the size of the array by the end of training.

We observed no positive evidence for transfer from digit span training to circle span, or vice versa. Neither was there strong evidence that digit span training benefited performance on the untrained verbal serial-recall tests of spoken digit span or letter span. The lack of transfer across verbal serial tasks is consistent with the predictions of the cognitive-routine framework (Gathercole et al., 2019 ). According to this theory, transfer occurs only when the demands of the training tasks cannot readily be met by existing STM mechanisms and processes. Only under these conditions will participants need to develop new cognitive routines that control and coordinate the processes involved in performing the task. When training involves only simple verbal serial-recall tasks, no new routines are required because a well-established and highly practiced set of mechanisms is already in place within verbal STM. There should therefore be no substantial transfer, as we indeed found. As was noted by Gathercole et al. ( 2019 ), the exception to this would be individuals with an underdeveloped verbal STM system. In children who do not yet rehearse, rehearsal training does indeed increase memory span (Broadley & MacDonald, 1993 ; Johnston, Johnson, & Gray, 1987 ).

The absence of transfer from either serial-recall training program to untrained serial-recall tasks therefore provides no evidence that the capacity of STM can be expanded with intensive training. This is particularly noteworthy for digit span, as the transfer tasks were distinguished only by a single feature—input modality for spoken digit span (visual to auditory), and semantic category for letter span (digits to letters). According to the current understanding of verbal STM, each of these three stimulus forms should be equally readily represented in verbal STM (Baddeley et al., 1984 ). If training had acted to increase STM capacity, the benefits of this extra capacity should therefore extend to both tasks. One possibility is that training on a single task allows participants to develop category-specific complex recoding strategies that reduce memory load by permit chunking of multi-item sequences (Ericsson et al., 1980 ; Martin & Fernberger, 1929 ). Although this hypothesis may explain the corresponding lack of transfer to letter span, it is not consistent with the corresponding absence of transfer to spoken digit span. With equal access to phonological storage for visual and auditory inputs, any beneficial effects of digit-specific encoding strategies would be expected to extend to digits presented in either modality.

Training-induced changes therefore appear to be tied to the semantic category and input modality of the memory items, as well as to paradigm. Why, then, should performance on the trained task improve at all if, as the absence of transfer suggests, the capacity of verbal STM is unchanged? The present data showed a relatively modest increase of 18% in digit span with training. This is considerably smaller that the gains observed in studies that had explicitly trained digit span strategies involving recoding (Ericsson & Staszewski, 1989 ; Kliegl et al., 1987 ; Reisberg et al., 1984 ; Yoon et al., 2018 ). Moreover, the gains that we found in digit span in the present study appear to be tied to the specific conjunction of the trained task features. One way of explaining this is that extensive training on a single task in which all parameters are fixed (e.g., perceptual, timing, and categorical) allows participants to fine-tune their task model. This could be conceived as a form of learning that takes place within the established system of verbal STM. If performance is finely tuned to all features of a single task, even superficial deviations from the trained task might be sufficient to render the model suboptimal. In this way, subtle changes within an existing system could be detectable in training effects when all task features are preserved, but not generalize to other variants of the same paradigm.

Training on the color change detection measure of visual STM generated both substantial on-task training gains and transfer to another task, in which participants detected changes in the orientation of lines in a multi-item array. To our knowledge, this is the first time that training-induced change has been demonstrated for static visual STM, a resource-limited memory system that has been extensively investigated in recent years (e.g., Alvarez & Cavanagh, 2004 ; Bays, Catalao, & Husain, 2009 ). There are several possible explanations for this outcome. Applying the rationale extended to findings of near transfer following WM training (Dahlin et al., 2008 ; Jaeggi et al., 2008 ; Klingberg, 2010 ), it could be interpreted as reflecting genuine plasticity in the capacity of visual STM. An alternative possibility is that the change detection paradigm may require the establishment of a new cognitive routine (Gathercole et al., 2019 ), and that this is the source of transfer. The very brief presentation of displays containing highly similar objects for a binary change detection judgment certainly imposes highly unfamiliar cognitive demands that quite plausibly might not be met solely through the processes in place within visual STM. The relatively large magnitude of the training gains seen in this task is certainly consistent with mediated learning. Perhaps, then, trainees develop a change detection routine to optimize their performance that—unlike the highly specific tuning to the specific task features seen in digit span training, which shows no substantial generalization—can be readily adapted to the untrained orientation detection task with its very similar demands.

Alternatively, transfer across change detection tasks reflects learning by the participant about the statistical properties of the displays. Such learning might underpin the robust training and transfer gains found for change detection tasks. Orhan and Jacobs ( 2014 ) have suggested that the apparent capacity limitations in visual STM might be due to a mismatch between the participant’s internal model and the true statistics of the stimuli. For example, our change detection tasks had a statistical structure of the elements within the display: Colors were not positioned at random but were constrained to have a minimal separation. This became even more constraining as the number of stimuli in the display increased. When the maximum number of items were in the display, they were closely packed, and there was much less room for variation in position than with fewer items. The orientation change task had the same statistical properties. Perhaps, then, learning about the statistics of the displays in one task could readily transfer to the other. An important question as yet unanswered is whether transfer to other change detection tasks would persist if the statistics of the displays changed between the trained and untrained activities.

Participants might also learn about the characteristics of their internal representations of stimuli in change detection tasks. To optimize the readout of information from memory, participants need to have an accurate model of the internal representation that will be produced by a particular input. In Bayesian terms, they need to develop an accurate generative model of the task. This form of learning or adaptation is likely to be tied to the low-level perceptual properties of the stimuli. In the case of serial recall with letters or digits, this might involve nothing more than fine-tuning, but for a completely novel task like change detection, more work might need to be done.

In summary, on-task performance improves after extensive practice with serial recall of visually presented digits. However, there is little evidence that this improvement confers any advantage to recall of visually presented letters, auditorily presented digits, or sequences of spatial locations. Changing either the stimulus domain, the presentation modality, or the category of the memory items eliminated the benefits of training. Digit span training does not substantially improve the capacity of verbal STM. In contrast, training on an unfamiliar color change detection task produces large gains in performance that transfer to a line orientation change detection task. The large improvement in change detection was unexpected, as change detection is often used to estimate core visual STM capacity. This might have been a consequence of learning how to perform a novel task, in much the same way as for more complex WM tasks, and also how to optimize performance by exploiting the statistical properties of the displays.

Abrahamse, E., Van Dijck, J.-P., Majerus, S., & Fias, W. (2014). Finding the answer in space: The mental whiteboard hypothesis on serial order in working memory. Frontiers in Human Neuroscience , 8 , 932. https://doi.org/10.3389/fnhum.2014.00932

Adams, J. W., & Hitch, G. J. (1997). Working memory and children’s mental addition. Journal of Experimental Child Psychology , 67 , 21–38. https://doi.org/10.1006/jecp.1997.2397

Article   PubMed   Google Scholar  

Alloway, T. P. (2007). Automated working memory assessment. San Antonio: Pearson.

Google Scholar  

Alloway, T. P., Gathercole, S. E., Kirkwood, H., & Elliott, J. (2008). Evaluating the validity of the automated working memory assessment. Educational Psychology , 28 , 725–734.

Article   Google Scholar  

Alloway, T. P., Gathercole, S. E., & Pickering, S. J. (2006). Verbal and visuospatial short-term and working memory in children: are they separable? Child Development , 77 , 1698–1716. https://doi.org/10.1111/j.1467-8624.2006.00968.x

Alvarez, G. A., & Cavanagh, P. (2004). The capacity of visual short-term memory is set both by visual information load and by number of objects. Psychological Science , 15 , 106–111. https://doi.org/10.1111/j.0963-7214.2004.01502006.x

Anderson, J. R. (1982). Acquisition of cognitive skill. Psychological Review , 89 , 369–406. https://doi.org/10.1037/0033-295X.89.4.369

Astle, D. E., Barnes, J. J., Baker, K., Colclough, G. L., & Woolrich, M. W. (2015). Cognitive training enhances intrinsic brain connectivity in childhood. Journal of Neuroscience , 35 , 6277–6283. https://doi.org/10.1523/JNEUROSCI.4517-14.2015

Awh, E., Barton, B., & Vogel, E. K. (2007). Visual working memory represents a fixed number of items regardless of complexity. Psychological Science , 18 , 622–628.

Baddeley, A. D., Gathercole, S. E., & Papagno, C. (1998). The phonological loop as a language learning device. Psychological Review , 105 , 158–173. https://doi.org/10.1037/0033-295X.105.1.158

Baddeley, A. D., Lewis, V., & Vallar, G. (1984). Exploring the articulatory loop. Quarterly Journal of Experimental Psychology , 36A , 233–252.

Bavelier, D., Green, C. S., Pouget, A., & Schrater, P. (2012). Brain plasticity through the life span: Learning to learn and action video games. Annual Review of Neuroscience , 35 , 391–416. https://doi.org/10.1146/annurev-neuro-060909-152832

Bayliss, D. M., Jarrold, C., Gunn, D. M., & Baddeley, A. D. (2003). The complexities of complex span: Explaining individual differences in working memory in children and adults. Journal of Experimental Psychology: General , 132 , 71–92. https://doi.org/10.1037/0096-3445.132.1.71

Bays, P. M., Catalao, R. F. G., & Husain, M. (2009). The precision of visual working memory is set by allocation of a shared resource. Journal of Vision , 9 (10), 7. https://doi.org/10.1167/9.10.7

Brehmer, Y., Westerberg, H., & Bäckman, L. (2012). Working-memory training in younger and older adults: Training gains, transfer, and maintenance. Frontiers in Human Neuroscience , 6 , 63. https://doi.org/10.3389/fnhum.2012.00063

Article   PubMed   PubMed Central   Google Scholar  

Broadley, I., & MacDonald, J. (1993). Teaching short term memory skills to children with Down syndrome. Down Syndrome Research and Practice , 1 , 56–62.

Buschkuehl, M., Jaeggi, S. M., Mueller, S. T., Shah, P., & Jonides, J. (2017). Training change detection leads to substantial task-specific improvement. Journal of Cognitive Enhancement , 1 , 419–433.

Chase, W. G., & Ericsson, K. A. (1981). Skilled memory. In J. R. Anderson (Ed.), Cognitive skills and their acquisition (pp. 141–189). Hillsdale: Erlbaum.

Chein, J. M., & Morrison, A. B. (2010). Expanding the mind’s workspace: training and transfer effects with a complex working memory span task. Psychonomic Bulletin & Review , 17 , 193–199. https://doi.org/10.3758/PBR.17.2.193

Cortese, S., Ferrin, M., Brandeis, D., Buitelaar, J., Daley, D., Dittmann, R. W., . . . Sonuga-Barke, E. J. S. (2015). Cognitive training for attention-deficit/hyperactivity disorder: Meta-analysis of clinical and neuropsychological outcomes from randomized controlled trials. Journal of the American Academy of Child and Adolescent Psychiatry , 54 , 164–174. https://doi.org/10.1016/j.jaac.2014.12.010

Cowan, N. (2001). The magical number 4 in short-term memory: A reconsideration of mental storage capacity. Behavioral and Brain Sciences , 24 , 87–114, disc. 114–185. https://doi.org/10.1017/S0140525X01003922

Cowan, N., Elliott, E. M., Saults, J. S., Morey, C. C., Mattox, S., Hismjatullina, A., & Conway, A. R. A. (2005). On the capacity of attention: Its estimation and its role in working memory and cognitive aptitudes. Cognitive Psychology , 51 , 42–100. https://doi.org/10.1016/j.cogpsych.2004.12.001

Dahlin, E., Neely, A. S., Larsson, A., Backman, L., & Nyberg, L. (2008). Transfer of learning after updating training mediated by the striatum. Science , 320 , 1510–1512. https://doi.org/10.1126/science.1155466

Darling, S., Della Sala, S., Logie, R. H., & Cantagallo, A. (2006). Neuropsychological evidence for separating components of visuo-spatial working memory. Journal of Neurology , 253 , 176–180.

Della Sala, S., Gray, C., Baddeley, A., Allamano, N., & Wilson, L. (1999). Pattern span: A tool for unwelding visuo-spatial memory. Neuropsychologia , 37 , 1189–1199.

Dunning, D. L., & Holmes, J. (2014). Does working memory training promote the use of strategies on untrained working memory tasks? Memory & Cognition , 42 , 854–862. https://doi.org/10.3758/s13421-014-0410-5

Engle, R. W., Carullo, J. J., & Collins, K. W. (1991). Individual differences in working memory for comprehension and following directions. Journal of Educational Research , 84 , 253–262.

Ericsson, K., Chase, W. G., & Faloon, S. (1980). Acquisition of a memory skill. Science , 208 , 1181–1182.

Ericsson, K. A., & Staszewski, J. J. (1989). Skilled memory and expertise: Mechanisms of exceptional performance. In D. Klahr & K. Kotovsky (Eds.), Complex information processing: The impact of Herbert A. Simon (Vol. 2, pp. 235–267). Hillsdale: Erlbaum.

Gaspar, J. G., Neider, M. B., Simons, D. J., McCarley, J. S., & Kramer, A. F. (2013). Change detection: Training and transfer. PLoS ONE , 8 , e67781. https://doi.org/10.1371/journal.pone.0067781

Gathercole, S. E. (2006). Nonword repetition and word learning: The nature of the relationship. Applied Psycholinguistics , 27 , 513–543.

Gathercole, S. E., Dunning, D. L., Holmes, J., & Norris, D. (2019). Working memory training involves learning new skills. Journal of Memory and Language , 105 , 19–42. https://doi.org/10.1016/j.jml.2018.10.003

Geary, D. C., Hoard, M. K., Byrd-Craven, J., & DeSoto, M. C. (2004). Strategy choices in simple and complex addition: contributions of working memory and counting knowledge for children with mathematical disability. Journal of Experimental Child Psychology , 88 , 121–151.

Gray, S., Chaban, P., Martinussen, R., Goldberg, R., Gotlieb, H., Kronitz, R., . . . Tannock, R. (2012). Effects of a computerized working memory training program on working memory, attention, and academics in adolescents with severe LD and comorbid ADHD: A randomized controlled trial. Journal of Child Psychology and Psychiatry , 53 , 1277–1284.

Hanley, J. R., Young, A. W., & Pearson, N. A. (1991). Impairment of the visuo-spatial sketch pad. Quarterly Journal of Experimental Psychology , 43A , 101–125. https://doi.org/10.1080/14640749108401001

Hardy, K. K., Willard, V. W., Allen, T. M., & Bonner, M. J. (2013). Working memory training in survivors of pediatric cancer: a randomized pilot study. Psycho-Oncology , 22 , 1856–1865. https://doi.org/10.1002/pon.3222

Harrison, T. L., Shipstead, Z., Hicks, K. L., Hambrick, D. Z., Redick, T. S., & Engle, R. W. (2013). Working memory training may increase working memory capacity but not fluid intelligence. Psychological Science , 24 , 2409–2419. https://doi.org/10.1177/0956797613492984

Henry, L. A., Messer, D. J., & Nash, G. (2014). Testing for near and far transfer effects with a short, face-to-face adaptive working memory training intervention in typical children. Infant and Child Development , 23 , 84–103.

Hurlstone, M. J., Hitch, G. J., & Baddeley, A. D. (2014). Memory for serial order across domains: An overview of the literature and directions for future research. Psychological Bulletin , 140 , 339–373. https://doi.org/10.1037/a0034221

Jaeggi, S. M., Buschkuehl, M., Jonides, J., & Perrig, W. J. (2008). Improving fluid intelligence with training on working memory. Proceedings of the National Academy of Sciences , 105 , 6829–6833. https://doi.org/10.1073/pnas.0801268105

Jaroslawska, A. J., Gathercole, S. E., Logie, M. R., & Holmes, J. (2016). Following instructions in a virtual school: Does working memory play a role? Memory & Cognition , 44 , 580–589.

JASP Team. (2015). JASP (Version 0.6.6) [Computer software]. Retrieved from https://jasp-stats.org/download/

Jeffreys, H. (1961). The theory of probability (3rd). Oxford: Oxford University Press, Clarendon Press.

Johnston, R. S., Johnson, C., & Gray, C. (1987). The emergence of the word length effect in young children: The effects of overt and covert rehearsal. British Journal of Developmental Psychology , 5 , 243–248.

Kliegl, R., Smith, J., Heckhausen, J., & Baltes, P. B. (1987). Mnemonic training for the acquisition of skilled digit memory. Cognition and Instruction , 4 , 203–223.

Klingberg, T. (2010). Training and plasticity of working memory. Trends in Cognitive Sciences , 14 , 317–324. https://doi.org/10.1016/j.tics.2010.05.002

Logie, R., & Pearson, D. G. (1997). The inner eye and the inner scribe of visuo-spatial working memory: Evidence from developmental fractionation. European Journal of Cognitive Psychology , 9 , 241–257.

Luck, S. J., & Vogel, E. K. (1997). The capacity of visual working memory for features and conjunctions. Nature , 390 , 279–281. https://doi.org/10.1038/36846

Majerus, S., D’Argembeau, A., Martinez Perez, T., Belayachi, S., Van der Linden, M., Collette, F., … Maquet, P. (2010). The commonality of neural networks for verbal and visual short-term memory. Journal of Cognitive Neuroscience , 22 , 2570–2593. https://doi.org/10.1162/jocn.2009.21378

Martin, P. R., & Fernberger, S. W. (1929). Improvement in memory span. American Journal of Psychology , 41 , 91–94.

McLean, J. F., & Hitch, G. J. (1999). Working memory impairments in children with specific arithmetic learning difficulties. Journal of Experimental Child Psychology , 74 , 240–260. https://doi.org/10.1006/jecp.1999.2516

Melby-Lervåg, M., & Hulme, C. (2013). Is working memory training effective? A meta-analytic review. Developmental Psychology , 49 , 270–291. https://doi.org/10.1037/a0028228

Melby-Lervåg, M., & Hulme, C. (2016). There is no convincing evidence that working memory training is effective: A reply to Au et al. (2014) and Karbach and Verhaeghen (2014). Psychonomic Bulletin & Review , 23 , 324–330.

Minear, M., Brasher, F., Guerrero, C. B., Brasher, M., Moore, A., & Sukeena, J. (2016). A simultaneous examination of two forms of working memory training: Evidence for near transfer only. Memory & Cognition , 44 , 1014–1037.

Norris, D. (2017). Short-term memory and long-term memory are still different. Psychological Bulletin , 143 , 992–1009.

Orhan, A. E., & Jacobs, R. A. (2014). Are performance limitations in visual short-term memory tasks due to capacity limitations or model mismatch? arXiv preprint. arXiv:1407.0644

Reisberg, D., Rappaport, I., & O’Shaughnessy, M. (1984). Limits of working memory: The digit digit-span. Journal of Experimental Psychology: Learning, Memory, and Cognition , 10 , 203–221. https://doi.org/10.1037/0278-7393.10.2.203

Simons, D. J., Boot, W. R., Charness, N., Gathercole, S. E., Chabris, C. F., Hambrick, D. Z., & Stine-Morrow, E. A. L. (2016). Do “brain-training” programs work? Psychological Science in the Public Interest , 17 , 103–186. https://doi.org/10.1177/1529100616661983

Staszewski, J. J. (1990). Exceptional memory: The influence of practice and knowledge on the development of elaborative encoding strategies. In Interactions among aptitudes, strategies, and knowledge in cognitive performance (pp. 252–285). New York: Springer.

Taatgen, N. A. (2013). The nature and transfer of cognitive skills. Psychological Review , 120 , 439–471. https://doi.org/10.1037/a0033138

Wechsler, D. (1981). WAIS–R manual. New York: Psychological Corp.

Xu, Z., Adam, K., Fang, X., & Vogel, E. (2018). The reliability and stability of visual working memory capacity. Behavior Research Methods , 50 , 576–588.

Yoon, J. S., Ericsson, K. A., & Donatelli, D. (2018). Effects of 30 years of disuse on exceptional memory performance. Cognitive Science , 42 (Suppl. 3), 884–903. https://doi.org/10.1111/cogs.12562

Download references

Author information

Authors and affiliations.

MRC Cognition and Brain Sciences Unit, University of Cambridge, Cambridge, UK

Dennis G. Norris, Jane Hall & Susan E. Gathercole

You can also search for this author in PubMed   Google Scholar

Corresponding author

Correspondence to Dennis G. Norris .

Additional information

Publisher’s note.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Electronic supplementary material

(DOCX 18 kb)

Rights and permissions

OpenAccess This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.

Reprints and permissions

About this article

Norris, D.G., Hall, J. & Gathercole, S.E. Can short-term memory be trained?. Mem Cogn 47 , 1012–1023 (2019). https://doi.org/10.3758/s13421-019-00901-z

Download citation

Published : 27 February 2019

Issue Date : 15 July 2019

DOI : https://doi.org/10.3758/s13421-019-00901-z

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Short-term memory
• Memory training
• Find a journal
• Publish with us
• Track your research

PERSPECTIVE article

A common short-term memory retrieval rate may describe many cognitive procedures.

• Department of Psychological Sciences, University of Missouri, Columbia, MO, USA

We examine the relationship between response speed and the number of items in short-term memory (STM) in four different paradigms and find evidence for a similar high-speed processing rate of about 25–30 items per second (∼35–40 ms/item). We propose that the similarity of the processing rates across paradigms reflects the operation of a very basic covert memory process, high-speed retrieval, that is involved in both the search for information in STM and the reactivation or refreshing of information that keeps it in STM. We link this process to a specific pattern of rhythmic, repetitive neural activity in the brain (gamma oscillations). This proposal generates ideas for research and calls for an integrative approach that combines neuroscientific measures with behavioral cognitive techniques.

An important feature of human information processing is short-term memory (STM), the ability to retain a small amount of information in a highly accessible state for a short time. The capacity of STM is limited to a certain number of items, and a key issue in cognitive psychology is the reason why STM is limited. Here we suggest that, over the last 40–50 years, at least four different paradigms have been developed that provide insights into the temporal properties of STM. Despite the wide variety of paradigms, we observed an intriguing similarity in a high-speed processing rate of about 25–30 items per second, which can be inferred from the relationship between response speed and memory load. We propose that the similarity of the processing rates across paradigms may reflect a basic covert memory process (i.e., a memory process that is inferred from the pattern of recall performance across certain conditions, rather than being directly observable), high-speed retrieval, which can be used for either recognition of a probe item or reactivation (refreshing) of an item for the sake of maintenance. We also link this process to recent developments in the neuroscientific literature and discuss implications for future research.

The Relationship Between Response Speed and Memory Load in Four Paradigms

After the seminal article of Miller (1956) on STM capacity limitations, human STM research mainly investigated the determinants of failure of STM by focusing on accuracy and error patterns in simple memory tasks. In the late 1960s, however, a complementary approach became increasingly popular. This approach consisted of studying how much time participants need in order to succeed in simple memory tasks. Specifically, Saul Sternberg studied how much time participants needed in order to indicate whether a probe item was present in a small set of memorized elements ( Sternberg, 1966 , 1969a ). The rationale was that, if the information in memory is needed to select the appropriate response, then the time taken to give that response will reveal something about the process by which one is searching in memory for that information. In order to explore the timing of memory search, Sternberg proposed what we term the Sternberg Item-Recognition paradigm (Figure 1A ). Although it is still the standard paradigm to investigate memory search rates, at least three other paradigms can be identified as providing insights into the temporal properties of STM (Figures 1B–D ); all show a positive relation between the number of items to be retained in STM (memory load) and the time it takes to respond to a probe item (response latency). Figure 2A provides an overview of what, based on our review, seem to be necessary boundary conditions that must be met to observe a clear positive relation between memory load and response latency. In what follows, only studies that met these conditions are reported and, when interpreting the observed common processing rate, we will explicitly address the role of these boundary conditions.

Figure 1. Schematic presentation of four paradigms providing insights in the relation between response speed and memory load. In the example, participants are presented with three letters to be maintained: K, B and N. The hand symbol together with the hourglass refers to a response given by pressing a button for which the speed is the variable of interest here. In (A) the Sternberg Item-Recognition paradigm, we examined speed of response to probe as a function of the number of memory items; in (B) through (D) we examined speed of response to processing items as a function of the concurrent number of items in memory; ( B) Brown-Peterson Pre-Load paradigm; (C) Complex Span paradigm; (D) Psychological Refractory Period paradigm.

Figure 2. (A) Schematic presentation of cognitive interpretation of the observed processing rate together with the boundary conditions (in black) that must be met to observe a clear relationship between memory load and response latency. Two different levels in STM are presented: a central level and a peripheral level. For verbal stimuli, the peripheral level offers an alternative maintenance mechanism (articulatory rehearsal), the use of which should be minimized when examining STM load for continuously presented items. (B) Estimates of STM retrieval slope for three kinds of verbal materials based on (1) the Sternberg Item-Recognition paradigm (gray bars), (2) the Brown-Peterson Pre-Load paradigm (red bar, second from the bottom), (3) the Complex Span paradigm (blue bar, last in the Words cluster), and (4) the Psychological Refractory Period (yellow bar, bottom). For the Sternberg Item-Recognition paradigm, the figure only includes studies that provided the information necessary to calculate 95% confidence intervals (represented by error bars). For the other paradigms, the unweighted average across studies mentioned in the text is presented.

For the sake of comparison, in the following review, we only included data of experiments that used simple verbal stimuli to be memorized (digits, letters and words), using healthy young adults as participants. We only included studies that provide the information necessary to examine a particular key index of the rate of retrieving information from STM, the slope of the function that relates response latency to STM load. Consequently, we only considered studies that included at least two different levels of memory load and that either reported the slope of the relation of interest or reported response latency for each memory load condition so that we could estimate the slope (averaged across positive and negative responses). Despite the fact that the four paradigms differ quite substantially in their methodology (see Figure 1 ), we identified a similar processing rate across them.

The Sternberg Item-Recognition Paradigm

The variable of interest is the speed with which participants decide whether the probe is a member of set of items held in STM by pressing, as quickly as possible without making errors (Figure 1A ). It is assumed that this decision requires people to scan through the content of STM to compare the probe with each item in memory. Delay of the response is interpreted as the operation of this time-consuming covert memory search. The classical finding is that response times increase linearly with the size of the memory set with a slope of about 35–40 ms per additional item in memory ( Sternberg, 1966 , 1969a ). The slope of this function is assumed to reflect the time it takes to retrieve a single item from STM. These classic findings of Sternberg launched a very productive line of investigation on memory search in cognitive science, with an overwhelming number of studies testing the original paradigm and variants of it. Because of the limited space here, the included studies using this paradigm were limited to the visual modality for presentation of both memory items and test items. On average, response latency increased at a rate of 37 ms per additional item held in memory. 1

The Brown-Peterson Pre-Load Paradigm

In the paradigm developed by Brown (1958) and Peterson and Peterson (1959) , a few stimuli to be remembered are followed by a processing task that is different enough to avoid material-specific interference, but challenging enough to prevent attention to the memoranda or rehearsal of them (Figure 1B ). The main finding was that memory is lost rapidly across about 30 s. The variable of interest here, though, is processing speed on the concurrent processing task that precedes recall. Slowing down has been shown in several studies comparing response speed under concurrent memory load with response speed without a concurrent load (e.g., Shulman and Greenberg, 1971 ; Baddeley and Hitch, 1974 ; Logan, 1978 ). It is assumed that, during the retention interval filled with processing, people engage in storage-related activities. When processing and storage both rely on attentional resources, storage-related activities are expected to postpone concurrent processing activities. Methodological details can be found in Footnote 2. 2 Vergauwe et al. (2014) found that response latency increased linearly at a rate of 43 ms per additional item held in memory.

The Complex Span Paradigm

In this paradigm, the presentation of items to be remembered is interleaved with items to be processed (Figure 1C ). The purpose was originally to assess the capability of working memory under the assumption that storage and processing share a common cognitive resource so that both of them must be engaged in order for capability to be assessed (e.g., Daneman and Carpenter, 1980 ). The variable of interest for the present purposes is processing speed on the concurrent processing task. Several studies have shown longer response latencies in later processing phases (high memory load), compared to the first processing phase (low memory load; e.g., Friedman and Miyake, 2004 ; Chen and Cowan, 2009 ). As for the Brown-Peterson pre-load paradigm, the underlying assumption is that slower processing reflects resource-sharing between attention-demanding processing and storage activities. Methodological details can be found in Footnote 2. Jarrold et al. (2011) found linear trends across the successive processing phases showing that response latency increases at an average rate of 37 ms per additional item held in memory (41 ms in Experiment 1 and 33 ms in Experiment 2).

Another potential variable of interest, but one that requires further work, is the time it takes to retrieve the next item to be recalled. Cowan (1992) measured the timing of spoken recall for simple digit span in children and proposed that each inter-word pause reflects a process of search through working memory to find the next digit to be recalled. Subsequent work ( Cowan et al., 1998 ) showed that the inter-word pauses for correctly-recalled lists did increase in approximately a linear fashion with increasing list length, in children in first grade (84 ms/item), third grade (58 ms/item), and fifth grade (25 ms/item). In adults, further work is needed to establish the scanning rate. One might worry that verbal rehearsal processes would play a role, though a relation between the spoken recall rate and search rates based on the scanning paradigm was demonstrated by Cowan et al. (1998) and by Hulme et al. (1999) . In complex span, presumably rehearsal processes have been interrupted by the processing task. Recall in these tasks, however, might involve more than a simple search, for example an attempt to use the processing task as a context to retrieve the list items. Thus, Cowan et al. (2003) noted that inter-word pauses in the responses lasted 4–10 times longer than in simple span.

Psychological Refractory Period Paradigm

This paradigm ( Welford, 1952 ; Pashler, 1994 ) usually combines two processing tasks requiring two responses in succession on a single trial. The original point was to explore processing demands by studying how the processing for the first response delayed the second response. In the task variants of interest here, memory demands are combined with processing demands. After the memory set is presented, at various stimulus-onset asynchronies (SOAs), a single stimulus pertaining to the processing task is presented, to which a speeded response is required (Figure 1D ). Some of these studies also manipulated the size of the memory set, which makes them of particular interest here. The finding of interest here is that the single speeded response took longer as more items were held in memory concurrently (e.g., Jolicoeur and Dell’Acqua, 1998 ; Stevanovski and Jolicoeur, 2007 ). Again, the underlying assumption is that processing and storage interfere with each other because they rely on a common attentional resource, resulting in slower processing. Methodological details can be found in Footnote 3. 3 Processing took about 46 ms longer per additional item in memory (32 and 60 ms in Stevanovski and Jolicoeur, 2007 , in Experiments 2 and 3, respectively).

Empirical Summary

We have identified a pattern that holds across four different paradigms: response speed slows down at a rate of about 30–40 ms per additional simple verbal item in memory (see Figure 2B ). The similarity across the paradigms suggests strongly the existence of a high-speed processing rate in STM of about 25–30 items per second (the equivalent of 40–33 ms/item).

Previous studies have pointed out the similarity between the processing rates observed in the complex span paradigm and the Sternberg item-recognition paradigm ( Jarrold et al., 2011 ), and between the rates observed in the Brown-Peterson pre-load paradigm and the Sternberg item-recognition paradigm ( Vergauwe et al., 2014 ). The present contribution is to note the similarity of processing rate across a wider range of procedures, and to propose a cognitive interpretation of this high-speed processing rate, in the next section.

Cognitive Interpretation of High-Speed Processing Rate in Human Short-Term Memory (STM)

We interpret the identified processing rate as reflecting the operation of a very basic covert memory process, retrieval from STM. In this view, although information retrieval and maintenance are typically referred to as different stages in STM, they are proposed to rely on the same process. When responding to a probe in the Sternberg task, high-speed retrieval is used in the service of memory search. It brings items in the focus of attention so that one can check whether it matches the probe. The slope observed in this task reflects directly the use of high-speed retrieval. In the three remaining paradigms, high-speed retrieval is used in the service of memory maintenance; it brings items in the focus of attention so that the information gets reactivated or refreshed. When high-speed retrieval and concurrent processing share a common resource (attention), the use of high-speed retrieval influences concurrent processing speed so that response latency increases for each additional item that is maintained. Under the assumption that maintenance is accomplished through sequential reactivation of information in a cumulative fashion, starting from the first list item and proceeding in forward order until the end, the observed rate reflects the rate at which items are reactivated in STM. In the Sternberg task, it is assumed that the presentation of the probe initiates a complete cycle through STM. In the other paradigms described here, storage is combined with a self-paced processing task and the idea is that a complete cycle of refreshing is interpolated before attention-demanding processing takes place. Thus, provided that participants aim at performing well on the memory task, attention is first used for a complete cycle through STM before it is shifted to the next processing stimulus. It is possible, though, that the same assumption might not hold in tasks in which the processing task is to be performed at a predefined pace (i.e., computer-paced). In these tasks, every processing item is typically followed by a variable period of free time during which refreshing can take place in a continuous matter. If the process of refreshing is exhaustive in nature, one might expect that, upon the presentation of the next processing item, on average only half of the items in STM would still need to be reactivated. Slopes relating response times to memory load would then reflect the amount of time it takes to scan half of the number of items in STM.

A schematic presentation of our cognitive interpretation of the observed processing rate is shown in Figure 2A . Two different levels in STM are presented: (1) a central level that is domain-general in nature, closely related to attention, and (2) a peripheral level that is domain-specific in nature and independent from the central level. High-speed retrieval is used at the central level to bring information into the focus of attention.

Together with the observation of Cowan et al. (1998) that retrieval rate as measured in a search task correlates with memory span, the identification of a rapid retrieval rate across several paradigms is directly relevant to the long-standing debate regarding the nature of the severe capacity limit of STM. Theoretically, the capacity limit of STM might reflect the number of items that can be active simultaneously within a given time-window. If one assumes that there is a limited time-window within which the items need to be reactivated so that all of them can be retained, then the capacity limit of STM would depend on the retrieval rate with faster rates resulting in more items reactivated within the fixed time-window. A similar idea was proposed by Cavanagh (1972) who showed an inverse relation between STM span and memory search rate for different materials. The speed of retrieval in STM also indicates that STM functions in a way that is much more rapid and dynamic than most people would think. Importantly, we consider this rapid retrieval rate to be independent of the slower verbal rehearsal rate that relies on covert speech, even though both might serve the same goal of maintaining information in STM (see Cowan et al., 1998 ; Hulme et al., 1999 ; Camos et al., 2011 ).

Note that although the idea of a limited time-window implies the existence of time-based forgetting in STM, it is not incompatible with interference-based forgetting. When items are not reactivated in time, forgetting might occur either because memory traces have decayed or because newer representations have overwritten previous ones or have become confusable with the previous ones. The degree of confusability might then depend on the number of features that are shared between the representations in STM. Moreover, Ricker and Cowan (2014) have recently shown that the process of consolidation influences the observed rate of forgetting over time with more consolidation leading to slower rates of time-based forgetting. This finding indicates that the relationship between STM capacity, retrieval rate and decay rate might depend on the robustness of the trace. Also, the length of the critical time-window might differ between individuals and this possibility needs to be taken into consideration when focusing on the relation between high-speed retrieval and STM capacity across individuals.

Boundary Conditions

There are studies in which the slope of the relationship between response speed and memory load was substantially smaller than the proposed constant of about 37 ms per item in normal adults. For example, in a Sternberg task, Banks and Atkinson (1974) forced participants to respond so quickly that they made a lot of errors. A flatter slope may occur when speed is stressed at the expense of accuracy because participants base their response on a feeling of familiarity, which can occur for all items in parallel, rather than on a more time-consuming but accurate item-by-item memory search. Burrows and Okada (1975) showed that the Sternberg slope changes at the limits of STM with a shallow slope of 13 ms when considering memory loads ranging between 8 and 20 words, for which the only viable mechanism might be familiarity. In the processing times within complex span, but using viewing or reading times rather than simple reaction times (RT) slopes across memory loads vary considerably (e.g., Engle et al., 1992 ; Friedman and Miyake, 2004 ). Viewing or reading might be covertly interrupted for refreshing, evading measurement. There are also studies in which the slope of the relationship between response speed and memory load was larger than the proposed constant of about 37 ms per item. RT slopes across memory loads are considerably steeper (up to about 100 ms per item) in studies that use Sternberg-like tasks in which participants need to have access to serial order information in order to judge the probe correctly, as opposed to the typical Sternberg task in which access to item information is sufficient (e.g., Sternberg, 1969b ; Ravizza et al., 2011 ; Majerus et al., 2012 ). Furthermore, studies in which a delay of several seconds was inserted between the presentation of the memory set and the presentation of the probe also reported somewhat steeper slopes (about 50–55 ms per item; e.g., Cairo et al., 2004 ; Chen and Desmond, 2005 ). Maintenance-related processes such as verbal rehearsal might take place during this delay and as such, influence the observed retrieval rate at the end of the trial. We suggest boundary conditions to observe a clear, positive relation between memory load and response latency, as presented in Figure 2A .

Relating High-Speed Retrieval in Short-Term Memory (STM) to Oscillations in the Brain

Recent neuroscientific developments lead to a view of retrieval rate as governed by oscillations (rhythmic, repetitive neural activity; e.g., Lisman and Jensen, 2013 ). In the dual oscillation model of STM ( Lisman and Idiart, 1995 ) it is proposed that the features of one item are active at the same time and are represented by a group of neurons that fire in the same gamma cycle (30–80 Hz). Next, the features of a second item are active at the same time and represented by the second gamma cycle within the same theta cycle (4–8 Hz). Lisman and Idiart (1995) linked the Sternberg slope to the duration of a gamma cycle. One item would be searched each time its gamma cycle of neural activity occurred. They also suggested that STM capacity limits could be determined by the number of gamma cycles that fit into one theta cycle. Given current uncertainties in these figures, this neural theory is reasonably compatible with a cognitive proposal by which STM capacity depends on the number of items that can be reactivated within a given time-window so that several items can be retained in a refreshed state simultaneously. Each gamma cycle would allow the refreshment of one item in STM. Our empirical retrieval rate of 37 ms/item would correspond to a gamma cycle of 27 Hz and would allow 3–6 items per theta cycle.

Thus, we propose to extend the view of Lisman and Idiart so that it encompasses our expanded function of high-speed retrieval. In this view, refreshing consists in the rapid reactivation of a limited number of items at a rate that reflects the length of one gamma cycle per item. In support of a link between STM maintenance and gamma oscillations, changes of oscillatory activity in the human gamma frequency band related to STM retention have been observed (e.g., Tallon-Baudry et al., 1998 ; Jokisch and Jensen, 2007 ; Meltzer et al., 2008 ) and Howard et al. (2003) showed that, in a Sternberg-type task, gamma power during retention was higher for larger memory sets. Furthermore, Roux et al. (2012) showed a relation between gamma-band activity and memory load in a left prefrontal area of the brain that has been associated with refreshing (e.g., Johnson et al., 2005 ). In this study, a number of red disks were displayed in different locations. After a short delay, a single red disk was shown and participants decided whether its location matched one of the study locations. An increase in gamma-band power between load 3 and load 6 was observed during the delay and this increase correlated with memory performance. Finally, Kamiński et al. (2011) found a negative correlation between individual’s STM performance and gamma cycle length. This is exactly the kind of relationship one would expect if STM capacity depends on the number of items that can be reactivated within a given time-window with each gamma cycle allowing the reactivation of one item in STM.

Conclusion and Outlook

The current proposal is novel in at least two ways. First, it proposes that the identified high-speed processing rate of about 27 items per second across four different procedures might reflect the operation of a very basic process of high-speed retrieval that serves both memory search and attention-based refreshing in STM. Thus, the attentional component of memory search and refreshing is proposed to be the same. This does not preclude the theoretical possibility that refreshing is equivalent to retrieval plus some additional operations; it only restricts these additional operations to a set of operations that do not require attention. Second, it proposes that this general process might be associated with gamma brain oscillations. We believe that our proposal has the potential of providing novel insights into the significant questions of how information is maintained in STM and why it’s capacity-limited. The proposal is based on a limited number of studies at this point and further research is needed, but the present proposal suggests several clear directions for further research.

Behavioral research should aim at testing the unique predictions that follow from our proposal. First, memory search and maintenance are proposed to rely on the same STM retrieval process. One direction is to look for interference patterns between both processes. The results of ongoing research of ours suggest that the memory search slope varies as a function of the time available to refresh memoranda. Another test of our proposal would be to compare the processing rates across the four paradigms in a within-participants design. This might also help us understand whether variations in the processing rate between procedures and materials are meaningful. Second, refreshing of a series of items is proposed to be enacted by consecutive gamma oscillations. Does the order of spontaneous refreshing follow the order of presentation? Can the distance between individual items in STM be described in terms of the number of gamma cycles that separate them? When a set of multiple items is successfully chunked into a few chunks, can we observe a decrease in the number of gamma cycles one needs to run through in order to refresh the entire set? In addition, future research should aim at testing the universality of the identified retrieval process by searching whether a similar processing rate can be observed in other paradigms and by examining response time distribution data. Another remaining question is whether transferring new external information into STM would occur at the same rate. The results of some studies suggest a slower rate of consolidation of about 200–250 ms per item in the Brown-Peterson pre-load paradigm (e.g., Jarrold et al., 2011 ; Vergauwe et al., 2014 ). This rate matches the length of theta cycles which have been linked to encoding new information (e.g., Klimesch, 1999 ). Finally, neurophysiological and cognitive approaches should be integrated to examine whether the length of gamma cycles and retrieval rate are influenced by the same factors (experimental, individual, developmental, clinical), and whether externally induced changes in gamma frequency (e.g., through magnetic stimulation of the brain) affect STM speed and capacity.

Conflict of Interest Statement

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Acknowledgments

This research was supported by Swiss National Science Foundation Grant PA00P1_139604 to Evie Vergauwe and NIH Grant R01-HD21338 to Nelson Cowan.

• ^ The average of 37 ms/item refers to an unweighted average across 8 slopes for digits ( M = 36 ms/item), 5 slopes for words ( M = 36 ms/item) and 13 slopes for letters ( M = 38 ms/item). When only taking into account the studies that provide information to calculate 95% confidence intervals (i.e., the studies included in the lower panel of Figure 2 ), the unweighted average is 36 ms/item.
• ^ In the Brown-Peterson pre-load paradigm and the complex span paradigm, a difference is typically made between the RT for the individual’s first response in the processing phase, referred to as first processing times and the mean of all subsequent RTs in that processing phase, referred to as subsequent processing times. While longer first processing times have been attributed to the consolidation of memory traces, longer subsequent processing times are typically attributed to the maintenance of memory traces (e.g., Engle et al., 1992 ; Jarrold et al., 2011 ). Because our focus is on the maintenance process rather than on consolidation, the current manuscript only reports analyses that concern subsequent processing times. The slopes reported for the complex span paradigm concern an average across different list lengths (from 4 to 7 words).
• ^ We included only conditions in which participants are required to recall all of the information after this speeded response because, only in those conditions, participants are required to make the processing response while keeping the information active in memory. Furthermore, we only included studies in which the effect of memory load on response time did not depend on SOA because only in those studies, the effect can be interpreted as a cost related to maintaining information in STM rather than consolidating information into STM.

Baddeley, A. D., and Hitch, G. (1974). “Recent advances in learning and motivation,” in Working Memory (Vol. 8), ed G. A. Bower (New York: Academic Press), 647–667.

Banks, W., and Atkinson, R. (1974). Accuracy and speed strategies in scanning active memory. Mem. Cognit. 2, 629–636. doi: 10.3758/bf03198131

Pubmed Abstract | Pubmed Full Text | CrossRef Full Text

Brown, H. L., and Kirsner, K. (1980). A within–subjects analysis of the relationship between memory span and processing rate in short–term memory. Cogn. Psychol. 12, 177–187.

Brown, J. (1958). Some tests of the decay theories of immediate memory. Q. J. Exp. Psychol. 10, 12–21. doi: 10.1080/17470215808416249

CrossRef Full Text

Burrows, D., and Okada, R. (1971). Serial position effects in high-speed memory search. Percept. Psychophys. 10, 305–308. doi: 10.3758/bf03212831

Burrows, D., and Okada, R. (1975). Memory retrieval from long and short lists. Science 188, 1031–1033. doi: 10.1126/science.188.4192.1031

Cairo, T. A., Liddle, P. F., Woodward, T. S., and Ngan, E. T. C. (2004). The influence of working memory load on phase specific patterns of cortical activity. Brain Res. Cogn. Brain Res. 21, 377–387. doi: 10.1016/j.cogbrainres.2004.06.014

Camos, V., Mora, G., and Oberauer, K. (2011). Adaptive choice between articulatory rehearsal and attentional refreshing in verbal working memory. Mem. Cognit. 39, 231–244. doi: 10.3758/s13421-010-0011-x

Cavanagh, J. P. (1972). Relation between the immediate memory span and the memory search rate. Psychol. Rev. 79, 525–530. doi: 10.1037/h0033482

Chen, S. H. A., and Desmond, J. E. (2005). Temporal dynamics of cerebro-cerebellar network recruitment during a cognitive task. Neuropsychologia 43, 1227–1237. doi: 10.1016/j.neuropsychologia.2004.12.015

Chen, Z., and Cowan, N. (2009). How verbal memory loads consume attention. Mem. Cognit. 37, 829–836. doi: 10.3758/mc.37.6.829

Clifton, C., and Tash, J. (1973). Effect of syllabic word length on memory-search rate. J. Exp. Psychol. 99, 231–235. doi: 10.1037/h0034643

Cowan, N. (1992). Verbal memory span and the timing of spoken recall. J. Mem. Lang. 31, 668–684.

Cowan, N., Towse, J. N., Hamilton, Z., Saults, J. S., Elliott, E. M., Lacey, J. F., et al. (2003). Children’s working-memory processes: a response-timing analysis. J. Exp. Psychol. Gen. 132, 113–132. doi: 10.1037/0096-3445.132.1.113

Cowan, N., Wood, N. L., Wood, P. K., Keller, T. A., Nugent, L. D., and Keller, C. V. (1998). Two separate verbal processing rates contributing to short-term memory span. J. Exp. Psychol. Gen. 127, 141–160. doi: 10.1037/0096-3445.127.2.141

Daneman, M., and Carpenter, P. A. (1980). Individual differences in working memory and reading. J. Verbal Learning Verbal Behav. 19, 450–466.

Engle, R. W., Cantor, J., and Carullo, J. J. (1992). Individual differences in working memory and comprehension: a test of four hypotheses. J. Exp. Psychol. Learn. Mem. Cogn. 18, 972–992. doi: 10.1037//0278-7393.18.5.972

Friedman, N. P., and Miyake, A. (2004). The reading span test and its predictive power for reading comprehension ability. J. Mem. Lang. 51, 136–158. doi: 10.1016/j.jml.2004.03.008

Howard, M. W., Rizzuto, D. S., Caplan, J. B., Madsen, J. R., Lisman, J., Aschenbrenner-Scheibe, R., et al. (2003). Gamma oscillations correlate with working memory load in humans. Cereb. Cortex 13, 1369–1374. doi: 10.1093/cercor/bhg084

Hulme, C., Newton, P., Cowan, N., Stuart, G., and Brown, G. (1999). Think before you speak: pauses, memory search and trace redintegration processes in verbal memory span. J. Exp. Psychol. Learn. Mem. Cogn. 25, 447–463. doi: 10.1037/0278-7393.25.2.447

Jarrold, C., Tam, H., Baddeley, A. D., and Harvey, C. E. (2011). How does processing affect storage in working memory tasks? Evidence for both domain-general and domain-specific effects. J. Exp. Psychol. Learn. Mem. Cogn. 37, 688–705. doi: 10.1037/a0022527

Johnson, M. K., Raye, C. L., Mitchell, K. J., Greene, E. J., Cunningham, W. A., and Sanislow, C. A. (2005). Using fMRI to investigate a component process of reflection: prefrontal correlates of refreshing a just-activated representation. Cogn. Affect. Behav. Neurosci. 5, 339–361. doi: 10.3758/CABN.5.3.339

Jokisch, D., and Jensen, O. (2007). Modulation of gamma and alpha activity during a working memory task engaging the dorsal or ventral stream. J. Neurosci. 27, 3244–3251. doi: 10.1523/jneurosci.5399-06.2007

Jolicoeur, P., and Dell’Acqua, R. (1998). The demonstration of short-term consolidation. Cogn. Psychol. 36, 138–202. doi: 10.1006/cogp.1998.0684

Kamiński, J., Brzezicka, A., and Wrobel, A. (2011). Short-term memory capacity (7 ± 2) predicted by theta to gamma cycle length ratio. Neurobiol. Learn. Mem. 95, 19–23. doi: 10.1016/j.nlm.2010.10.001

Klimesch, W. (1999). EEG alpha and theta oscillations reflect cognitive and memory performance: a review and analysis. Brain Res. Brain Res. Rev. 29, 169–195. doi: 10.1016/s0165-0173(98)00056-3

Lisman, J. E., and Idiart, M. A. (1995). Storage of 7 +/−2 short–term memories in oscillatory subcycles. Science 267, 1512–1515. doi: 10.1126/science.7878473

Lisman, J. E., and Jensen, O. (2013). The theta-gamma neural code. Neuron 77, 1002–1016. doi: 10.1016/j.neuron.2013.03.007

Logan, G. D. (1978). Attention in character classification: evidence for the automaticity of component stages. J. Exp. Psychol. Gen. 107, 32–63. doi: 10.1037/0096-3445.107.1.32

Majerus, S., Attout, L., D’Argembeau, A., Degueldre, C., Fias, W., Maquet, P., et al. (2012). Attention supports verbal short-term memory via competition between dorsal and ventral attention networks. Cereb. Cortex 22, 1086–1097. doi: 10.1093/cercor/bhr174

Meltzer, J. A., Zaveri, H. P., Goncharova, I. I., Distasio, M. M., Papademetris, X., Spencer, S. S., et al. (2008). Effects of working memory load on oscillatory power in human intracranial EEG. Cereb. Cortex 18, 1843–1855. doi: 10.1093/cercor/bhm213

Miller, G. A. (1956). The magical number seven, plus or minus two: some limits on our capacity for processing information. Psychol. Rev. 63, 81–97. doi: 10.1037/h0043158

Pashler, H. (1994). Dual–task interference in simple tasks: data and theory. Psychol. Bull. 116, 220–244. doi: 10.1037//0033-2909.116.2.220

Peterson, L. R., and Peterson, M. J. (1959). Short-term retention of individual verbal items. J. Exp. Psychol. 58, 193–198. doi: 10.1037/h0049234

Puckett, J. M., and Kausler, D. H. (1984). Individual differences and models of memory span: a role for memory search rate? J. Exp. Psychol. Learn. Mem. Cogn. 10, 72–82. doi: 10.1037/0278-7393.10.1.72

Ravizza, S. M., Hazeltine, E., Ruiz, S., and Zhu, D. C. (2011). Left TPJ activity in verbal working memory: implications for storage- and senory-specific models of short term memory. Neuroimage 55, 1836–1846. doi: 10.1016/j.neuroimage.2010.12.021

Ricker, T. J., and Cowan, N. (2014). Differences in presentations methods in working memory procedures: a matter of working memory consolidation. J. Exp. Psychol. Learn. Mem. Cogn. 40, 417–428. doi: 10.1037/a0034301

Roux, F., Wibral, M., Mohr, H. M., Singer, W., and Uhlhaas, P. J. (2012). Gamma-band activity in human prefrontal cortex codes for the number of relevant items maintained in working memory. J. Neurosci. 32, 12411–12420. doi: 10.1523/jneurosci.0421-12.2012

Shulman, H. G., and Greenberg, S. N. (1971). Perceptual deficit due to division of attention between memory and perception. J. Exp. Psychol. 88, 171–176. doi: 10.1037/h0030879

Sternberg, S. (1966). High speed scanning in human memory. Science 153, 652–654. doi: 10.1126/science.153.3736.652

Sternberg, S. (1969a). Memory scanning: mental processes revealed by reaction-time experiments. Am. Sci. 4, 421–457.

Pubmed Abstract | Pubmed Full Text

Sternberg, S. (1969b). Two invariances in retrieval of contextual information from memory. Paper presented at the Eastern Psychological Association meeting, April 1969. Also, Bell Telephone Laboratories Memorandum MM69 1221–1229.

Stevanovski, B., and Jolicoeur, P. (2007). Visual short-term memory: central capacity limitations in short-term consolidation. Vis. Cogn. 15, 532–563. doi: 10.1080/13506280600871917

Tallon-Baudry, C., Bertrand, O., Peronnet, F., and Pernier, J. (1998). Induced gamma-band activity during the delay of a visual short-term memory task in humans. J. Neurosci. 18, 4244–4254.

Vergauwe, E., Camos, V., and Barrouillet, P. (2014). The effect of storage on processing: how is information maintained in working memory? J. Exp. Psychol. Learn. Mem. Cogn. doi: 10.1037/a0035779

CrossRef Full Text . [Epub ahead of print].

Welford, A. T. (1952). An apparatus for use in studying serial performance. Am. J. Psychol. 65, 91–97. doi: 10.2307/1418834

Keywords: short-term memory, working memory, attention, retrieval, refreshing, memory search

Citation: Vergauwe E and Cowan N (2014) A common short-term memory retrieval rate may describe many cognitive procedures. Front. Hum. Neurosci. 8 :126. doi: 10.3389/fnhum.2014.00126

Received: 10 December 2013; Accepted: 19 February 2014; Published online: 07 March 2014

Reviewed by:

Copyright © 2014 Vergauwe and Cowan. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY) . The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

*Correspondence: Nelson Cowan, Department of Psychological Sciences, University of Missouri, 210 McAlester Hall, Columbia, MO 65211, USA e-mail: [email protected]

Disclaimer: All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher.

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

• View all journals
• Explore content
• About the journal
• Publish with us
• Sign up for alerts

Short-term memory articles within Scientific Reports

Article 17 June 2024 | Open Access

The influence of Chinese typography on information dissemination in graphic design: based on eye-tracking data

• Weilong Chen
• , Jiqiang Yang
•  &  Yiluo Wang

Article 14 June 2024 | Open Access

Evidence of object permanence, short-term spatial memory, causality, understanding of object properties and gravity across five different ungulate species

• Alina Schaffer
• , Anja Widdig
•  &  Federica Amici

Article 07 March 2024 | Open Access

Be prepared for interruptions: EEG correlates of anticipation when dealing with task interruptions and the role of aging

• , Stephan Getzmann
•  &  Daniel Schneider

Article 22 January 2024 | Open Access

Oculomotor inhibition markers of working memory load

• Oren Kadosh
• , Kfir Inbal
•  &  Yoram S. Bonneh

Article 27 November 2023 | Open Access

Proprioceptive short-term memory in passive motor learning

• Shinya Chiyohara
• , Jun-ichiro Furukawa
•  &  Hiroshi Imamizu

Article 20 November 2023 | Open Access

Delaying circadian sleep phase under ultradian light cycle causes time-of-day-dependent alteration of cognition and mood

• Fanny Fuchs
• , Ludivine Robin-Choteau
•  &  Patrice Bourgin

Article 07 September 2023 | Open Access

A quantitative analysis of spontaneous alternation behaviors on a Y-maze reveals adverse effects of acute social isolation on spatial working memory

• , Hyeyeon Kang
•  &  Doyun Lee

Article 02 September 2023 | Open Access

Rats adaptively seek information to accommodate a lack of information

• , Yoshio Sakurai
•  &  Dai Yanagihara

Article 24 December 2022 | Open Access

Impairment in novelty-promoted memory via behavioral tagging and capture before apparent memory loss in a knock-in model of Alzheimer’s disease

• Tabitha Broadbelt
• , Menekse Mutlu-Smith
•  &  Szu-Han Wang

Article 25 October 2022 | Open Access

Task-dependent fractal patterns of information processing in working memory

• Jeremi K. Ochab
• , Marcin Wątorek
•  &  Paweł Oświęcimka

Article 23 September 2022 | Open Access

Navigating the impact of workplace distractions for persons with TBI: a qualitative descriptive study

• DeAnna Pinnow
• , Renee Causey-Upton
•  &  Peter Meulenbroek

Article 17 May 2022 | Open Access

Developmental differences in the impact of perceptual salience on short-term memory performance and meta-memory skills

• Tiziana Pedale
• , Serena Mastroberardino
•  &  Valerio Santangelo

Article 07 January 2022 | Open Access

Working memory and pattern separation in founder strains of the BXD recombinant inbred mouse panel

• Price E. Dickson
•  &  Guy Mittleman

Article 17 September 2021 | Open Access

Cognitive control affects motor learning through local variations in GABA within the primary motor cortex

• Shuki Maruyama
• , Masaki Fukunaga
•  &  Norihiro Sadato

Article 18 June 2021 | Open Access

Neonatal administration of a subanaesthetic dose of JM-1232(−) in mice results in no behavioural deficits in adulthood

• Koji Iwanaga
• , Yasushi Satoh
•  &  Takehiko Ikeda

Article 13 May 2021 | Open Access

Comparing physiological responses during cognitive tests in virtual environments vs. in identical real-world environments

• Saleh Kalantari
• , James D. Rounds
•  &  Jesus G. Cruz-Garza

Article 22 April 2021 | Open Access

Delayed-matching-to-position working memory in mice relies on NMDA-receptors in prefrontal pyramidal cells

• Kasyoka Kilonzo
• , Bastiaan van der Veen
•  &  Dennis Kätzel

Article 09 February 2021 | Open Access

TDCS effects on pointing task learning in young and old adults

• E. Kaminski
• , M. Engelhardt
•  &  P. Ragert

Article 27 October 2020 | Open Access

Effects of long-lasting social isolation and re-socialization on cognitive performance and brain activity: a longitudinal study in Octodon degus

• Daniela S. Rivera
• , Carolina B. Lindsay
•  &  Nibaldo C. Inestrosa

Article 09 October 2020 | Open Access

Association between free-living sleep and memory and attention in healthy adolescents

• Runa Stefansdottir
• , Hilde Gundersen
•  &  Erlingur Johannsson

Article 16 April 2020 | Open Access

Possible association between spindle frequency and reversal-learning in aged family dogs

• Ivaylo Borislavov Iotchev
• , Dóra Szabó
•  &  Enikő Kubinyi

Article 22 November 2019 | Open Access

Unsupervised machine-learning classification of electrophysiologically active electrodes during human cognitive task performance

• Krishnakant V. Saboo
• , Yogatheesan Varatharajah
•  &  Michal T. Kucewicz

Article 25 October 2019 | Open Access

Five-Year-Old Children’s Working Memory Can Be Improved When Children Act On A Transparent Goal Cue

• Christophe Fitamen
• , Agnès Blaye
•  &  Valérie Camos

Article 30 September 2019 | Open Access

M1 muscarinic receptor is a key target of neuroprotection, neuroregeneration and memory recovery by i-Extract from Withania somnifera

• Arpita Konar
• , Richa Gupta
•  &  Mahendra K. Thakur

Article 15 July 2019 | Open Access

Effects of Auditory Distraction on Face Memory

• , Laura Mieth
•  &  Axel Buchner

Article 03 July 2019 | Open Access

Elevation of endocannabinoids in the brain by synthetic cannabinoid JWH-018: mechanism and effect on learning and memory

• , Ryo Fukumori
•  &  Yuji Ishii

Article 04 February 2019 | Open Access

The Effect of Body-Related Stimuli on Mental Rotation in Children, Young and Elderly Adults

• Tina Iachini
• , Gennaro Ruggiero
•  &  Francesco Ruotolo

Article 25 June 2018 | Open Access

Molecular diversity of clustered protocadherin-α required for sensory integration and short-term memory in mice

• Tatsuya Yamagishi
• , Kohei Yoshitake
•  &  Katsuei Shibuki

Article 30 April 2018 | Open Access

TMS Over the Cerebellum Interferes with Short-term Memory of Visual Sequences

• , Z. Cattaneo
•  &  T. Vecchi

Article 14 February 2018 | Open Access

Electrophysiological evidence of memory-based detection of auditory regularity violations in anesthetized mice

• Jari L. O. Kurkela
• , Arto Lipponen
•  &  Piia Astikainen

Article 28 July 2017 | Open Access

Neuropeptide Y prolongs non-social memory and differentially affects acquisition, consolidation, and retrieval of non-social and social memory in male mice

• Johannes Kornhuber
•  &  Iulia Zoicas

Article 21 June 2017 | Open Access

Differential regulations of vestibulo-ocular reflex and optokinetic response by β- and α2-adrenergic receptors in the cerebellar flocculus

• , Soshi Tanabe
•  &  Tomoo Hirano

Article 24 October 2016 | Open Access

Retro-dimension-cue benefit in visual working memory

• Chaoxiong Ye
• , Zhonghua Hu
•  &  Qiang Liu

Article 20 June 2016 | Open Access

Newly-formed emotional memories guide selective attention processes: Evidence from event-related potentials

• Harald T. Schupp
• , Ursula Kirmse
•  &  Britta Renner

Article 10 November 2015 | Open Access

APOE-ε4 selectively modulates posteromedial cortex activity during scene perception and short-term memory in young healthy adults

• J. P. Shine
• , C. J. Hodgetts
•  &  K. S. Graham

Article 15 October 2015 | Open Access

Distinct roles of the RasGAP family proteins in C. elegans associative learning and memory

• M. Dávid Gyurkó
• , Péter Csermely
•  &  Attila Steták

Article 08 January 2013 | Open Access

Increased NR2A:NR2B ratio compresses long-term depression range and constrains long-term memory

• Zhenzhong Cui
• , Ruiben Feng
•  &  Joe Z. Tsien

Browse broader subjects

• Learning and memory

Quick links

• Explore articles by subject
• Guide to authors
• Editorial policies

• Search Menu
• Sign in through your institution
• CNS Injury and Stroke
• Epilepsy and Sleep
• Movement Disorders
• Multiple Sclerosis/Neuroinflammation
• Neuro-oncology
• Neurodegeneration - Cellular & Molecular
• Neuromuscular Disease
• Neuropsychiatry
• Pain and Headache
• Advance articles
• Editor's Choice
• Author Guidelines
• Submission Site
• Why publish with this journal?
• Open Access
• About Brain
• Editorial Board
• Advertising and Corporate Services
• Journals Career Network
• Self-Archiving Policy
• Dispatch Dates
• Terms and Conditions
• Journals on Oxford Academic
• Books on Oxford Academic

Article Contents

The human hippocampus contributes to short-term memory.

• Article contents
• Figures & tables
• Supplementary Data

Masud Husain, The human hippocampus contributes to short-term memory, Brain , 2024;, awae194, https://doi.org/10.1093/brain/awae194

• Permissions Icon Permissions

Dogma can have unhealthy consequences. Open any major textbook—on neuroscience, psychology or neurology—and you will find in its pages the conventional view that the human hippocampus plays a crucial role in long-term memory (LTM), also referred to as episodic memory. The evidence for this claim is now very well established. But for some eminent authors, such as Squire and his colleagues, 1 it comes with an equally important caveat: namely, that the hippocampus does not contribute to short-term memory (STM). The two parts of this proposal, concerning the differential role of the hippocampus in LTM and STM, are now so engrained in many researchers’ minds that it can often be difficult to challenge. It has become dogma and, much more than that, an obstacle to progress.

The proposal has had a profound impact on both our fundamental understanding of memory systems, as well as application to clinical syndromes. In neuroscience, this perspective on the role of the hippocampus is now an established and important pillar in the argument that STM and LTM systems are segregated, functionally and anatomically. In clinical application, this dogma has meant that tests of hippocampal function and integrity invariably rely on measuring LTM. On the other hand, STM has been considered to be the preserve of frontoparietal networks. However, intracranial recording studies in humans, including an important recent report, 2 adds to a growing body of evidence which supports the view that the hippocampus does indeed have a role to play in STM.

One of the first challenges to the idea that the contribution of the hippocampus is confined to LTM came with two reports that amnesic patients with bilateral lesions involving this brain region suffered from an impairment of STM. Olson and colleagues 3 assessed recall using visual tasks that required participants to retain information over only 1 or 8 s. They found that although their patients could remember three objects or three locations, they were highly impaired when asked to report objects as well as their locations. Thus, although single features (objects or locations) could be retained by hippocampal patients, there was a clear deficit for retention of feature conjunctions (object-location bindings). Similar conclusions were reached by Hannula et al. 4 who showed that their amnesic hippocampal patients were impaired at retaining either object-in-scenes or faces-in-scenes, even when the probes were presented immediately after participants had viewed the study arrays.

Subsequently, other investigations reported that such amnesic patients are also impaired when asked to reproduce the locations of objects that they had viewed on a table after very short delays. 5 In fact, these hippocampal cases made a specific type of error, which was termed a ‘swap error’—swapping the locations of items that they had seen. In other words, they recalled the objects and locations that they had observed, but when asked to reproduce these, they misbound objects with their locations. Crucially, these were not making random guesses. Furthermore, hippocampal amnesics were significantly impaired at detecting changes in object locations in scenes compared to detecting changes in objects alone, even when probed immediately after viewing a scene, 6 or when tested after a 5-s retention interval on a visual match-to-sample task. 7

One important criticism of these studies though is that the hippocampal lesions were not necessarily focal. The patients tested had suffered either prolonged anoxia 3-6 (which might have widespread effects beyond hippocampal volume loss visible on neuroimaging), encephalitis 3 , 5 , 6 (usually herpes simplex), surgical resections of lesions causing epilepsy 7 or even closed head injury, 6 with some individuals clearly having extensive medial temporal lobe (MTL) damage well beyond the hippocampus. A further argument by Squire and colleagues 1 that has been levelled against these findings is that the effects observed are actually best interpreted as an impairment of LTM. In their view, these patients actually have intact STM, so they can perform well when the number of items that have to be retained is small (below the capacity of STM). However, when STM capacity limit is breached and because they no longer can rely on using LTM due to their hippocampal lesions, an impairment becomes evident, for example on tasks where they have to reproduce the locations of several objects. 1 , 5

Although such objections might seem reasonable, data from limbic encephalitis patients 8 with very focal MTL lesions involving the hippocampus 9 have been invaluable in evaluating these claims. Using a task that tests the ability to remember the identities and locations of fractals presented on a touchscreen, Pertzov and colleagues 8 demonstrated that these individuals had good recollection for single items but they started to make significant object-location misbinding errors when there were only three fractals and their positions required to be maintained in memory, over just 4 s. Three items would presumably be within conventional STM capacity limits, so no breach of STM should have occurred and yet the patients were significantly impaired (see also the findings from Olson et al. 3 ). Furthermore, the type of errors that these limbic encephalitis and other hippocampal lesion patients made were not random guesses, as might be expected if the issue was simply one of breaching STM capacity limits. Instead, the major impairment was in binding the fractal objects to their locations, echoing previous findings, 3 , 5 , 7 including from Squire’s own laboratory. 1

To add to these reports of individuals with hippocampal pathology, several investigations have now shown activity of hippocampal neurons related to encoding, maintenance and retrieval on STM tasks. 10 The latest of these is a formidable study of human intracranial recordings from 36 patients who were being assessed for possible surgical treatment for drug-resistant epilepsy. 2 These participants performed a Sternberg task, viewing either one or three pictures, each shown for 2 s. After a delay of 2.8 s or less, they were presented with a probe picture and asked to indicate whether it was one of the ones they had seen previously in that trial. In total, recordings were obtained from 1454 single neurons and 1922 microware channels for local field potentials, across several brain regions that were sampled.

In the hippocampus, the activity of so-called ‘category’ cells (those that showed preferential firing every time an image belonging to a particular visual category, e.g. faces or animals or cars, was presented) was modulated by STM load (number of items in the trial) during the maintenance period. Their firing rate declined when the memory load increased from one to three items. Further, the spiking of these neurons was more synchronized with gamma frequency local field potentials when their preferred visual category had to be retained in memory. The investigators went on to show that in some hippocampal neurons, the phase of slow, theta waves (3–7 Hz) and the amplitude of faster gamma waves (30–140 Hz) was coupled—so called theta-gamma phase-amplitude coupling (PAC). Intriguingly, the extent to which this coupling was coordinated with theta activity in the medial frontal cortex was significantly greater in the high memory load (three items) compared to the low load (one item) condition. In other words, the timing of activity of these hippocampal neurons was coordinated with frontal theta activity, specifically when the need for more cognitive control increased (with greater STM load). The authors propose that such theta-gamma PAC might be an important mechanism for inter-regional control whereby frontal systems might modulate storage in the hippocampus. Regardless of whether this hypothesis is correct, the data clearly reveal hippocampal involvement and modulation of activity with memory load in all phases of a simple STM task.

From this wealth of findings, across lesion and human intracranial recordings, it is difficult to escape the conclusion that the human hippocampus does indeed contribute to STM. The highly specific misbinding effects that have been observed following hippocampal damage are not easy to sweep away either as epiphenomena or the result of some artificial effect that results from the hippocampus normally helping out when STM capacity is breached. Indeed, they resonate extremely closely with views that propose that the hippocampus has a special role in relational binding—bringing together the who, what, when and where—in long term, episodic memory. Thus, this brain region may have an important role in binding elements that belong to a memory, regardless of the duration of retention, long or short. This emerging view of hippocampal functions also has important implications for testing its integrity. The data very strongly suggest that this should be possible using brief tests of STM, and that long retention delays do not have to be deployed to observe the effects of hippocampal dysfunction. Dogma sometimes needs to be challenged. Otherwise, it can prevent progress.

Jeneson   A , Squire   LR . Working memory, long-term memory, and medial temporal lobe function . Learn Mem . 2012 ; 19 : 15 – 25 .

Google Scholar

Daume   J , Kamiński   J , Schjetnan   AGP , et al.    Control of working memory by phase–amplitude coupling of human hippocampal neurons . Nature . 2024 ; 629 : 393 – 401 .

Olson   IR , Page   K , Moore   KS , Chatterjee   A , Verfaellie   M . Working memory for conjunctions relies on the medial temporal lobe . J Neurosci . 2006 ; 26 : 4596 – 4601 .

Hannula   DE , Tranel   D , Cohen   NJ . The long and the short of it: Relational memory impairments in amnesia, even at short lags . J Neurosci . 2006 ; 26 : 8352 – 8359 .

Watson   PD , Voss   JL , Warren   DE , Tranel   D , Cohen   NJ . Spatial reconstruction by patients with hippocampal damage is dominated by relational memory errors . Hippocampus . 2013 ; 23 : 570 – 580 .

Hannula   DE , Tranel   D , Allen   JS , Kirchhoff   BA , Nickel   AE , Cohen   NJ . Memory for items and relationships among items embedded in realistic scenes: Disproportionate relational memory impairments in amnesia . Neuropsychology . 2015 ; 29 : 126 – 138 .

Finke   C , Braun   M , Ostendorf   F , et al.    The human hippocampal formation mediates short-term memory of colour-location associations . Neuropsychologia . 2008 ; 46 : 614 – 623 .

Pertzov   Y , Miller   TD , Gorgoraptis   N , et al.    Binding deficits in memory following medial temporal lobe damage in patients with voltage-gated potassium channel complex antibody-associated limbic encephalitis . Brain . 2013 ; 136 ( Pt 8 ): 2474 – 2485 .

Miller   TD , Chong   TT-J , Aimola Davies   AM , et al.    Focal CA3 hippocampal subfield atrophy following LGI1 VGKC-complex antibody limbic encephalitis . Brain . 2017 ; 140 : 1212 – 1219 .

Rutishauser   U , Reddy   L , Mormann   F , Sarnthein   J . The architecture of human memory: Insights from human single-neuron recordings . J Neurosci . 2021 ; 41 : 883 – 890 .

Email alerts

Citing articles via, looking for your next opportunity.

• Contact the editorial office
• Guarantors of Brain
• Recommend to your Library

Affiliations

• Online ISSN 1460-2156
• Print ISSN 0006-8950
• Copyright © 2024 Guarantors of Brain
• About Oxford Academic
• Publish journals with us
• University press partners
• What we publish
• New features  
• Open access
• Institutional account management
• Rights and permissions
• Get help with access
• Accessibility
• Advertising
• Media enquiries
• Oxford University Press
• Oxford Languages
• University of Oxford

Oxford University Press is a department of the University of Oxford. It furthers the University's objective of excellence in research, scholarship, and education by publishing worldwide

• Copyright © 2024 Oxford University Press
• Cookie settings
• Cookie policy
• Privacy policy
• Legal notice

This Feature Is Available To Subscribers Only

Sign In or Create an Account

This PDF is available to Subscribers Only

For full access to this pdf, sign in to an existing account, or purchase an annual subscription.

An official website of the United States government

The .gov means it’s official. Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

The site is secure. The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

• Publications
• Account settings

Preview improvements coming to the PMC website in October 2024. Learn More or Try it out now .

• Advanced Search
• Journal List
• HHS Author Manuscripts

What are the differences between long-term, short-term, and working memory?

Nelson cowan.

Department of Psychological Sciences, University of Missouri, 18 McAlester Hall, Columbia, MO 65211, USA

In the recent literature there has been considerable confusion about the three types of memory: long-term, short-term, and working memory. This chapter strives to reduce that confusion and makes up-to-date assessments of these types of memory. Long- and short-term memory could differ in two fundamental ways, with only short-term memory demonstrating (1) temporal decay and (2) chunk capacity limits. Both properties of short-term memory are still controversial but the current literature is rather encouraging regarding the existence of both decay and capacity limits. Working memory has been conceived and defined in three different, slightly discrepant ways: as short-term memory applied to cognitive tasks, as a multi-component system that holds and manipulates information in short-term memory, and as the use of attention to manage short-term memory. Regardless of the definition, there are some measures of memory in the short term that seem routine and do not correlate well with cognitive aptitudes and other measures (those usually identified with the term “working memory”) that seem more attention demanding and do correlate well with these aptitudes. The evidence is evaluated and placed within a theoretical framework depicted in Fig. 1 .

A depiction of the theoretical modeling framework. Modified from Cowan (1988) and refined in further work by Cowan (1995 , 1999 , 2005) .

Historical roots of a basic scientific question

How many phases of a memory are there? In a naïve view of memory, it could be made all of one cloth. Some people have a good ability to capture facts and events in memory, whereas others have less such ability. Yet, long before there were true psychological laboratories, a more careful observation must have shown that there are separable aspects of memory. An elderly teacher might be seen relating old lessons as vividly as he ever did, and yet it might be evident that his ability to capture the names of new students, or to recall which student made what comment in an ongoing conversation, has diminished over the years.

The scientific study of memory is usually traced back to Hermann Ebbinghaus (1885/1913 translation) , who examined his own acquisition and forgetting of new information in the form of series of nonsense syllables tested at various periods upto 31 days. Among many important observations, Ebbinghaus noticed that he often had a “first fleeting grasp … of the series in moments of special concentration” (p. 33) but that this immediate memory did not ensure that the series had been memorized in a way that would allow its recall later on. Stable memorization sometimes required further repetitions of the series. Soon afterward, James (1890) proposed a distinction between primary memory, the small amount of information held as the trailing edge of the conscious present, and secondary memory, the vast body of knowledge stored over a lifetime. The primary memory of James is like the first fleeting grasp of Ebbinghaus.

The Industrial Revolution made some new demands on what James (1890) called primary memory. In the 1850s, telegraph operators had to remember and interpret rapid series of dots and dashes conveyed acoustically. In 1876, the telephone was invented. Three years later, operators in Lowell, Massachusetts started using telephone numbers for more than 200 subscribers so that substitute operators could be more easily trained if the town’s four regular operators succumbed to a raging measles epidemic. This use of telephone numbers, complemented by a word prefix, of course spread. (The author’s telephone number in 1957 was Whitehall 2–6742; the number is still assigned, albeit as a seven-digit number.) Even before the book by Ebbinghaus, Nipher (1878) reported on the serial position curve obtained among the digits in logarithms that he tried to recall. The nonsense syllables that Ebbinghaus had invented as a tool can be seen to have acquired more ecological validity in an industrial age with expanding information demands, perhaps highlighting the practical importance of primary memory in daily life. Primary memory seems taxed as one is asked to keep in mind aspects of an unfamiliar situation, such as names, places, things, and ideas that one has not encountered before.

Yet, the subjective experience of a difference between primary and secondary memory does not automatically guarantee that these types of memory separately contribute to the science of remembering. Researchers from a different perspective have long hoped that they could write a single equation, or a single set of principles at least, that would capture all of memory, from the very immediate to the very long-term. McGeoch (1932) illustrated that forgetting over time was not simply a matter of an inevitable decay of memory but rather of interference during the retention interval; one could find situations in which memory improved, rather than diminish, over time. From this perspective, one might view what appeared to be forgetting from primary memory as the profound effect of interference from other items on memory for any one item, with interference effects continuing forever but not totally destroying a given memory. This perspective has been maintained and developed over the years by a steady line of researchers believing in the unity of memory, including, among others, Melton (1963) , Bjork and Whitten (1974) , Wickelgren (1974) , Crowder (1982 , 1993) , Glenberg and Swanson (1986) , Brown et al. (2000) , Nairne (2002) , Neath and Surprenant (2003) , and Lewandowsky et al. (2004) .

Description of three kinds of memory

In this chapter I will assess the strength of evidence for three types of memory: long-term memory, short-term memory, and working memory. Long-term memory is a vast store of knowledge and a record of prior events, and it exists according to all theoretical views; it would be difficult to deny that each normal person has at his or her command a rich, although not flawless or complete, set of long-term memories.

Short-term memory is related to the primary memory of James (1890) and is a term that Broadbent (1958) and Atkinson and Shiffrin (1968) used in slightly different ways. Like Atkinson and Shiffrin, I take it to reflect faculties of the human mind that can hold a limited amount of information in a very accessible state temporarily. One difference between the term “short-term memory” and the term “primary memory” is that the latter might be considered to be more restricted. It is possible that not every temporarily accessible idea is, or even was, in conscious awareness. For example, by this conception, if you are speaking to a person with a foreign accent and inadvertently alter your speech to match the foreign speaker’s accent, you are influenced by what was until that point an unconscious (and therefore uncontrollable) aspect of your short-term memory. One might relate short-term memory to a pattern of neural firing that represents a particular idea and one might consider the idea to be in short-term memory only when the firing pattern, or cell assembly, is active ( Hebb, 1949 ). The individual might or might not be aware of the idea during that period of activation.

Working memory is not completely distinct from short-term memory. It is a term that was used by Miller et al. (1960) to refer to memory as it is used to plan and carry out behavior. One relies on working memory to retain the partial results while solving an arithmetic problem without paper, to combine the premises in a lengthy rhetorical argument, or to bake a cake without making the unfortunate mistake of adding the same ingredient twice. (Your working memory would have been more heavily taxed while reading the previous sentence if I had saved the phrase “one relies on working memory” until the end of the sentence, which I did in within my first draft of that sentence; working memory thus affects good writing.) The term “working memory” became much more dominant in the field after Baddeley and Hitch (1974) demonstrated that a single module could not account for all kinds of temporary memory. Their thinking led to an influential model ( Baddeley, 1986 ) in which verbal-phonological and visual-spatial representations were held separately, and were managed and manipulated with the help of attention-related processes, termed the central executive. In the 1974 paper, this central executive possibly had its own memory that crossed domains of representation. By 1986, this general memory had been eliminated from the model, but it was added back again by Baddeley (2000) in the form of an episodic buffer . That seemed necessary to explain short-term memory of features that did not match the other stores (particularly semantic information in memory) and to explain cross-domain associations in working memory, such as the retention of links between names and faces. Because of the work of Baddeley et al. (1975) , working memory is generally viewed as the combination of multiple components working together. Some even include in that bundle the heavy contribution of long-term memory, which reduces the working memory load by organizing and grouping information in working memory into a smaller number of units ( Miller, 1956 ; Ericsson and Kintsch, 1995 ). For example, the letter series IRSCIAFBI can be remembered much more easily as a series of acronyms for three federal agencies of the United States of America: the Internal Revenue Service (IRS), the Central Intelligence Agency (CIA), and the Federal Bureau of Investigation (FBI). However, that factor was not emphasized in the well-known model of Baddeley (1986) .

What is clear from my definition is that working memory includes short-term memory and other processing mechanisms that help to make use of short-term memory. This definition is different from the one used by some other researchers (e.g., Engle, 2002 ), who would like to reserve the term working memory to refer only to the attention-related aspects of short-term memory. This, however, is not so much a debate about substance, but rather a slightly confusing discrepancy in the usage of terms.

One reason to pursue the term working memory is that measures of working memory have been found to correlate with intellectual aptitudes (and especially fluid intelligence) better than measures of short-term memory and, in fact, possibly better than measures of any other particular psychological process (e.g., Daneman and Carpenter, 1980 ; Kyllonen and Christal, 1990 ; Daneman and Merikle, 1996 ; Engle et al., 1999 ; Conway et al., 2005 ). It has been thought that this reflects the use of measures that incorporate not only storage but also processing, the notion being that both storage and processing have to be engaged concurrently to assess working memory capacity in a way that is related to cognitive aptitude. More recently, Engle et al. (1999) introduced the notion that aptitudes and working memory both depend on the ability to control attention, or to apply the control of attention to the management of both primary and secondary memory ( Unsworth and Engle, 2007 ). However, more research is needed on exactly what we learn from the high correlation between working memory and intellectual aptitudes, and this issue will be discussed further after the more basic issue of the short-term versus the long-term memory distinction is addressed.

Meanwhile, it may be helpful to summarize a theoretical framework ( Cowan, 1988 , 1995 , 1999 , 2001 , 2005 ) based on past research. This framework, illustrated in Fig. 1 , helps to account for the relation between long-term, short-term, and working memory mechanisms and explains what I see as the relation between them. In this framework, short-term memory is derived from a temporarily activated subset of information in long-term memory. This activated subset may decay as a function of time unless it is refreshed, although the evidence for decay is still tentative at best. A subset of the activated information is the focus of attention, which appears to be limited in chunk capacity (how many separate items can be included at once). New associations between activated elements can form the focus of attention. Now the evidence related to this modeling framework will be discussed.

The short-term memory/long-term memory distinction

If there is a difference between short- and long-term memory stores, there are two possible ways in which these stores may differ: in duration , and in capacity . A duration difference means that items in short-term storage decay from this sort of storage as a function of time. A capacity difference means that there is a limit in how many items short-term storage can hold. If there is only a limit in capacity, a number of items smaller than the capacity limit could remain in short-term storage until they are replaced by other items. Both types of limit are controversial. Therefore, in order to assess the usefulness of the short-term storage concept, duration and capacity limits will be assessed in turn.

Duration limits

The concept of short-term memory limited by decay over time was present even at the beginning of cognitive psychology, for example in the work of Broadbent (1958) . If decay were the only principle affecting performance in an immediate memory experiment, it would perhaps be easy to detect this decay. However, even in Broadbent’s work contaminating variables were recognized. To assess decay one must take into account, or overcome, contaminating effects of rehearsal, long-term retrieval, and temporal distinctiveness, which will be discussed one at a time in conjunction with evidence for and against decay.

Overcoming contamination from rehearsal

According to various researchers there is a process whereby one imagines how the words on the list are pronounced without saying them aloud, a process called covert verbal rehearsal. With practice, this process comes to occur with a minimum of attention. Guttentag (1984) used a secondary task to show that rehearsal of a list to be recalled was effortful in young children, but not in adults. If, in a particular experimental procedure, no loss of short-term memory is observed, one can attribute that response pattern to rehearsal. Therefore, steps have been taken to eliminate rehearsal through a process termed articulatory suppression, in which a simple utterance such as the word “ the” is repeatedly pronounced by the participant during part or all of the short-term memory task (e.g., Baddeley et al., 1975 ). There is still the possible objection that whatever utterance is used to suppress rehearsal unfortunately causes interference, which could be the true reason for memory loss over time instead of decay.

That problem of interference would appear moot in light of the findings of Lewandowsky et al. (2004) . They presented lists of letters to be recalled and varied how long the participant was supposed to take to recall each item in the list. In some conditions, they added articulatory suppression to prevent rehearsal. Despite that suppression, they observed no difference in performance with the time between items in the response varying between 400 and 1600 ms (or between conditions in which the word “super” was pronounced one, two, or three times between consecutive items in the response). They found no evidence of memory decay.

A limitation of this finding, though, is that covert verbal rehearsal may not be the only type of rehearsal that participants can use. Perhaps there are types that are not prevented by articulatory suppression. In particular, Cowan (1992) suggested that the process of mentally attending to words or searching through the list, an attention-demanding process, could serve to reactivate items to be recalled in a manner similar to covert verbal rehearsal. The key difference is that it would not be expected that articulatory suppression would prevent that type of rehearsal. Instead, to prevent that type of rehearsal an attention-demanding task would have to be used.

Barrouillet et al. (2004 , 2007 ) have results that do seem to suggest that there is another, more attention-demanding type of rehearsal. They have interposed materials between items to be recalled that require choices; they can be numbers to read aloud or multi-choice reaction times. It is found that these interfere with retention to an extent commensurate to the proportion of the inter-item interval used up attending to the distracting items. As the rate of the distracting items goes up, fewer of the to-be-recalled items are recalled. The notion is that when the distracting task does not require attention, the freed-up attention allows an attention- based rehearsal of the items to be recalled. When the interposed task is more automatic and does not require as much attention (e.g., an articulatory suppression task) there is much less effect of the rate of these interposed items.

Based on this logic, one could imagine a version of Lewandowsky’s task in which not articulatory suppression but attention-demanding verbal stimuli are placed between items in the response, and in which the duration of this filled time between items in the response varies from trial to trial. The verbal, attention-demanding stimuli should prevent both attention-based rehearsal and articulation-based rehearsal. If there is decay, then performance should decline across serial positions more severely when longer filled intervals are placed between items in the response. Unfortunately, though, such results might be accounted for alternatively as the result of interference from the distracting stimuli, without the need to invoke decay.

What seems to be needed, then, is a procedure to prevent both articulation-based and attention-based rehearsal without introducing interference. Cowan and Aubuchon (in press) tried out one type of procedure that may accomplish this. They presented lists of seven printed digits in which the time between items varied within a list. In addition to some randomly timed filler lists, there were four critical trial types, in which the six inter-digit blank intervals were all short (0.5 s following each item) or all long (2 s following each item), or comprised three short and then three long intervals, or three long and then three short intervals. Moreover, there were two post-list response cues. According to one cue, the participant was to recall the list with the items in the presented order, but at any rate they wished. According to the other response cue, the list was to be recalled using the same timing in which it was presented. The expectation was that the need to remember the timing in the latter response condition would prevent rehearsal of either type. As a consequence, performance should be impaired on trials in which the first three response intervals are long because, on these trials, there is more time for forgetting of most of the list items. Just as predicted, there was a significant interaction between the response cue and the length of the first half of the response intervals. When participants were free to recall items at their own pace, performance was no better with a short first half ( M =.71) than with a long first half ( M =.74). The slight benefit of a long first half in that situation could occur because it allowed the list to be rehearsed early on in the response. In contrast, when the timing of recall had to match the timing of the list presentation, performance was better with a short first half ( M = .70) than with a long first half ( M = .67). This, then, suggests there could be decay in short-term memory.

Overcoming contamination from long-term retrieval

If there is more than one type of memory storage then there still is the problem of which store provided the information underlying a response. There is no guarantee that, just because a procedure is considered a test of short-term storage, the long-term store will not be used. For example, in a simple digit span task, a series of digits is presented and is to be repeated immediately afterward from memory. If that series turned out to be only slightly different from the participant’s telephone number, the participant might be able to memorize the new number quickly and repeat it from long-term memory. The dual-store theories of memory allow this. Although Broadbent (1958) and Atkinson and Shiffrin (1968) drew their models of information processing as a series of boxes representing different memory stores, with long-term memory following short-term memory, these boxes do not imply that memory is exclusively in one box or another; they are better interpreted as the relative times of the first entry of information from a stimulus into one store and then the next. The question remains, then as to how one can determine if a response comes from short-term memory.

Waugh and Norman (1965) developed a mathematical model to accomplish this. The model operated with the assumption that long-term memory occurs for the entire list, including a plateau in the middle of the list. In contrast, by the time of recall, short-term memory is said to remain only at the end of the list. This model assumes that, for any particular serial position within a list, the likelihood of successful short-term storage (S) and long-term storage (L) are independent, so that the likelihood of recalling the item is S+L−SL.

A slightly different assumption is that short- and long-term stores are not independent but are used in a complementary fashion. The availability of short-term memory of an item may allow resources needed for long-term memorization to be shifted to elsewhere in the list. The data seem more consistent with that assumption. In several studies, lists to be recalled have been presented to patients with Korsakoff’s amnesia and normal control participants ( Baddeley and Warrington, 1970 ; Carlesimo et al., 1995 ). These studies show that, in immediate recall, performance in amnesic individuals is preserved at the last few serial positions of the list. It is as if the performance in those serial positions is based mostly or entirely on short-term storage, and that there is no decrease in that kind of storage in the amnesic patients. In delayed recall, the amnesic patients show a deficit at all serial positions, as one would expect if short-term memory for the end of the list is lost as a function of a filled delay period ( Glanzer and Cunitz, 1966 ).

Overcoming contamination from temporal distinctiveness

Last, it has been argued that the loss of memory over time is not necessarily the result of decay. Instead, it can be caused by temporal distinctiveness in retrieval. This kind of theory assumes that the temporal context of an item serves as a retrieval cue for that item, even in free recall. An item separated in time from all other items is relatively distinctive and easy to recall, whereas an item that is relatively close to other items is more difficult to recall because it shares their temporal cues to retrieval. Shortly after a list is presented the most recent items are the most distinct temporally (much like the distinctness of a telephone pole you are practically touching compared to poles extending further down the road). Across a retention interval, the relative distinctiveness of the most recent items decreases (much like standing far away from even the last pole in a series).

Although there are data that can be interpreted according to distinctiveness, there also are what look like dissociations between the effects of distinctiveness and a genuine short-term memory effect. One can see this, for example, in the classic procedure of Peterson and Peterson (1959) in which letter trigrams are to be recalled immediately or only after a distracting task, counting backward from a starting number by three, for a period lasting up to 18 s. Peterson and Peterson found severe memory loss for the letter trigram as the filled delay was increased. However, subsequently, sceptics argued that the memory loss occurred because the temporal distinctiveness of the current letter trigram diminished as the filled delay increased. In particular, this delay effect was said to occur because of the increase across test delays in the proactive interference from previous trials. On the first few trials, the delay does not matter ( Keppel and Underwood, 1962 ) and no detrimental effect of delay is observed if delays of 5, 10, 15, and 20 s are tested in separate trial blocks ( Turvey et al., 1970 ; Greene, 1996 ).

Yet, there may be a true decay effect at shorter test intervals. Baddeley and Scott (1971) set up a trailer in a shopping mall so that they could test a large number of participants for one trial each, so as to avoid proactive interference. They found an effect of the test delay within the first 5 s but not at longer delays. Still, it seems that the concept of decay is not yet on very firm ground and warrants further study. It may be that decay actually reflects not a gradual degradation of the quality of the short-term memory trace, but a sudden collapse at a point that varies from trial to trial. With a control for temporal distinctiveness, Cowan et al. (1997a) found what could be a sudden collapse in the representation of memory for a tone with delays of between 5 and 10 s.

Chunk capacity limits

The concept of capacity limits was raised several times in the history of cognitive psychology. Miller (1956) famously discussed the “magical number seven plus or minus two” as a constant in short-term processing, including list recall, absolute judgment, and numerical estimation experiments. However, his autobiographical essay ( Miller, 1989 ) indicates that he was never very serious about the number seven; it was a rhetorical device that he used to tie together the otherwise unrelated strands of his research for a talk. Although it is true that memory span is approximately seven items in adults, there is no guarantee that each item is a separate entity. Perhaps the most important point of Miller’s (1956) article was that multiple items can be combined into a larger, meaningful unit. Later studies suggested that the limit in capacity is more typically only three or four units ( Broadbent, 1975 ; Cowan, 2001 ). That conclusion was based on an attempt to take into account strategies that often increase the efficiency of use of a limited capacity, or that allow the maintenance of additional information separate from that limited capacity. To understand these methods of discussing capacity limits I will again mention three types of contamination. These come from chunking and the use of long-term memory, from rehearsal, and from non-capacity-limited types of storage.

Overcoming contamination from chunking and the use of long-term memory

A participant’s response in an immediate-memory task depends on how the information to be recalled is grouped to form multi-item chunks ( Miller, 1956 ). Because it is not usually clear what chunks have been used in recall, it is not clear how many chunks can be retained and whether the number is truly fixed. Broadbent (1975) proposed some situations in which multi-item chunk formation was not a factor, and suggested on the basis of results from such procedures that the true capacity limit is three items (each serving as a single-item chunk). For example, although memory span is often about seven items, errors are made with seven-item lists and the error-free limit is typically three items. When people must recall items from a category in long-term memory, such as the states of the United States, they do so in spurts of about three items on average. It is as if the bucket of short-term memory is filled from the well of long-term memory and must be emptied before it is refilled. Cowan (2001) noted other such situations in which multi-item chunks cannot be formed. For example, in running memory span, a long list of items is presented with an unpredictable endpoint, making grouping impossible. When the list ends, the participant is to recall a certain number of items from the end of the list. Typically, people can recall three or four items from the end of the list, although the exact number depends on task demands ( Bunting et al., 2006 ). Individuals differ in capacity, which ranges from about two to six items in adults (and fewer in children), and the individual capacity limit is a strong correlate of cognitive aptitude.

Another way to take into account the role of multi-item chunk formation is to set up the task in a manner that allows chunks to be observed. Tulving and Patkau (1962) studied free recall of word lists with various levels of structure, ranging from random words to well-formed English sentences, with several different levels of coherence in between. A chunk was defined as a series of words reproduced by the participant in the same order in which the words had been presented. It was estimated that, in all conditions, participants recalled an average of four to six chunks. Cowan et al. (2004) tried to refine that method by testing serial recall of eight-word lists, which were composed of four pairs of words that previously had been associated with various levels of learning (0, 1, 2, or 4 prior word–word pairings). Each word used in the list was presented an equal number of times (four, except in a non-studied control condition) but what varied was how many of those presentations were as singletons and how many were as a consistent pairing. The number of paired prior exposures was held constant across the four pairs in a list. A mathematical model was used to estimate the proportion of recalled pairs that could be attributed to the learned association (i.e., to a two-word chunk) as opposed to separate recall of the two words in a pair. This model suggested that the capacity limit was about 3.5 chunks in every learning condition, but that the ratio of two-word chunks to one-word chunks increased as a function of the number of prior exposures to the pairs in the list.

The issue of rehearsal is not entirely separate from the issue of chunk formation. In the traditional concept of rehearsal (e.g., Baddeley, 1986 ), one imagines that the items are covertly articulated in the presented order at an even pace. However, another possibility is that rehearsal involves the use of articulatory processes in order to put the items into groups. In fact, Cowan et al. (2006a) asked participants in a digit span experiment how they carried out the task and by far the most common answer among adults was that they grouped the items; participants rarely mentioned saying the items to themselves. Yet, it is clear that suppressing rehearsal affects performance.

Presumably, the situations in which items cannot be rehearsed are for the most part the same as the situations in which items cannot be grouped. For example, Cowan et al. (2005) relied on a running memory span procedure in which the items were presented at the rapid rate of 4 per second. At that rate, it is impossible to rehearse the items as they are presented. Instead, the task is probably accomplished by retaining a passive store (sensory or phonological memory) and then transferring the last few items from that store into a more attention-related store at the time of recall. In fact, with a fast presentation rate in running span, instructions to rehearse the items is detrimental, not helpful, to performance ( Hockey, 1973 ). Another example is memory for lists that were ignored at the time of their presentation ( Cowan et al., 1999 ). In these cases, the capacity limit is close to the three or four items suggested by Broadbent (1975) and Cowan (2001) .

It is still quite possible that there is a speech-based short-term storage mechanism that is by and large independent of the chunk-based mechanism. In terms of the popular model of Baddeley (2000) , the former is the phonological loop and the latter, the episodic buffer. In terms of Cowan (1988 , 1995 , 1999 , 2005) , the former is part of activated memory, which may have a time limit due to decay, and the latter is the focus of attention, which is assumed to have a chunk capacity limit.

Chen and Cowan (2005) showed that the time limit and chunk capacity limit in short-term memory are separate. They repeated the procedure of Cowan et al. (2004) in which pairs of words sometimes were presented in a training session preceding the list recall test. They combined lists composed of pairs as in that study. Now, however, both free and serial recall tasks were used, and the length of list varied. For long lists and free recall, the chunk capacity limit governed the recall. For example, lists of six well-learned pairs were recalled as well as lists of six unpaired singletons (i.e., were recalled at similar proportions of words correct). For shorter lists and serial recall strictly scored, the time limit instead governed the recall. For example, lists of four well-learned pairs were not recalled nearly as well as lists of four unpaired singletons, but only as well as lists of eight unpaired singletons. For intermediate conditions it appeared as if chunk capacity limits and time limits operate together to govern recall. Perhaps the capacity-limited mechanism holds items and the rehearsal mechanism preserves some serial order memory for those held items. The exact way in which these limits work together is not yet clear.

Overcoming contamination from non-capacity-limited types of storage

It is difficult to demonstrate a true capacity limit that is related to attention if, as I believe, there are other types of short-term memory mechanisms that complicate the results. A general capacity should include chunks of information of all sorts: for example, information derived from both acoustic and visual stimuli, and from both verbal and nonverbal stimuli. If this is the case, there should be cross-interference between one type of memory load and another. However, the literature often has shown that there is much more interference between similar types of memoranda, such as two visual arrays of objects or two acoustically presented word lists, than there is between two dissimilar types, such as one visual array and one verbal list. Cocchini et al. (2002) suggested that there is little or no interference between dissimilar lists. If so, that would appear to provide an argument against the presence of a general, cross-domain, short-term memory store.

Morey and Cowan (2004 , 2005) questioned this conclusion. They presented a visual array of colored spots to be compared to a second array that matched the first or differed from it in one spot’s color. Before the first array or just after it, participants sometimes heard a list of digits that was then to be recited between the two arrays. In a low-load condition, the list was their own seven-digit telephone number whereas, in a high-load condition, it was a random seven-digit number. Only the latter condition interfered with array-comparison performance, and then only if the list was to be recited aloud between the arrays. This suggests that retrieving seven random digits in a way that also engages rehearsal processes relies upon some type of short-term memory mechanism that also is needed for the visual arrays. That shared mechanism may be the focus of attention, with its capacity limit. Apparently, though, if the list was maintained silently rather than being recited aloud, this silent maintenance occurred without much use of the common, attention-based storage mechanism, so visual array performance was not much affected.

The types of short-term memory whose contribution to recall may obscure the capacity limit can include any types of activated memory that fall outside of the focus of attention. In the modeling framework depicted in Fig. 1 , this can include sensory memory features as well as semantic features. Sperling (1960) famously illustrated the difference between unlimited sensory memory and capacity-limited categorical memory. If an array of characters was followed by a partial report cue shortly after the array, most of the characters in the cued row could be recalled. If the cue was delayed about 1 s, most of the sensory information had decayed and performance was limited to about four characters, regardless of the size of the array. Based on this study, the four-character limit could be seen as either a limit in the capacity of short-term memory or a limit in the rate with which information could be transferred from sensory memory into a categorical form before it decayed. However, Darwin et al. (1972) carried out an analogous auditory experiment and found a limit of about four items even though the observed decay period for sensory memory was about 4s. Given the striking differences between Sperling and Darwin et al. in the time period available for the transfer of information to a categorical form, the common four-item limit is best viewed as a capacity limit rather than a rate limit.

Saults and Cowan (2007) tested this conceptual framework in a series of experiments in which arrays were presented in two modalities at once or, in another procedure, one after the other. A visual array of colored spots was supplemented by an array of spoken digits occurring in four separate loudspeakers, each one consistently assigned to a different voice to ease perception. On some trials, participants knew that they were responsible for both modalities at once whereas, in other trials, participants knew that they were responsible for only the visual or only the acoustic stimuli. They received a probe array that was the same as the previous array (or the same as one modality in that previous array) or differed from the previous array in the identity of one stimulus. The task was to determine if there was a change. The use of cross-modality, capacity-limited storage predicts a particular pattern of results. It predicts that performance on either modality should be diminished in the dual-modality condition compared to the unimodal conditions, due to strain on the cross-modality store. That is how the results turned out. Moreover, if the cross-modality, capacity-limited store were the only type of storage used, then the sum of visual and auditory capacities in the dual-modality condition should be no greater than the larger of the two unimodal capacities (which happened to be the visual capacity). The reason is that the limited-capacity store would hold the same number of units no matter whether they were all from one modality or were from two modalities combined. That prediction was confirmed, but only if there was a post-perceptual mask in both modalities at once following the array to be remembered. The post-perceptual mask included a multicolored spot at each visual object location and a sound composed of all possible digits overlaid, from each loudspeaker. It was presented long enough after the arrays to be recalled that their perception would have been complete (e.g., 1 s afterward; cf. Vogel et al., 2006 ). Presumably, the mask was capable of overwriting various types of sensory-specific features in activated memory, leaving behind only the more generic, categorical information present in the focus of attention, which presumably is protected from masking interference by the attention process. The limit of the focus of attention was again shown to be between three and four items, for either unimodal visual or bimodal stimuli.

Even without using masking stimuli, it may be possible to find a phase of the short-term memory process that is general across domains. Cowan and Morey (2007) presented two stimulus sets to be recalled (or, in control conditions, only one set). The two stimulus sets could include two spoken lists of digits, two spatial arrays of colored spots, or one of each, in either order. Following this presentation, a cue indicated that the participant would be responsible for only the first array, only the second array, or both arrays. Three seconds followed before a probe. The effect of memory load could be compared in two ways. Performance on those trials in which two sets of stimuli were presented and both were cued for retention could be compared either to trials in which only one set was presented, or it could be compared to trials in which both sets were presented but the cue later indicated that only one set had to be retained. The part of working memory preceding the cue showed modality-specific dual-task effects: encoding a stimulus set of one type was hurt more by also encoding another set if both sets were in the same modality. However, the retention of information following the cue showed dual-task effects that were not modality-specific. When two sets had been presented, retaining both of them was detrimental compared to retaining only one set (as specified by the post-stimulus retention cue to retain one versus both sets), and this dual-task effect was similar in magnitude no matter whether the sets were in the same or different modalities. After the initial encoding, working memory storage across several seconds thus may occur abstractly, in the focus of attention.

Other evidence for a separate short-term storage

Last, there is other evidence that does not directly support either temporal decay or a capacity limit specifically, but implies that one or the other of these limits exist. Bjork and Whitten (1974) and Tzeng (1973) made temporal distinctiveness arguments on the basis of what is called continual distractor list recall, in which a recency effect persists even when the list is followed by a distracter-filled delay before recall. The filled delay should have destroyed short-term memory but the recency effect occurs anyway, provided that the items in the list also are separated by distracter-filled delays to increase their distinctiveness from one another. In favor of short-term storage, though, other studies have shown dissociations between what is found in ordinary immediate recall versus continual distractor recall (e.g., word length effects reversed in continual distractor recall: Cowan et al., 1997b ; proactive interference at the most recent list positions in continual distractor recall only: Craik & Birtwistle, 1971 ; Davelaar et al., 2005 ).

There is also additional neuroimaging evidence for short-term storage. Talmi et al. (2005) found that recognition of earlier portions of a list, but not the last few items, activated areas within the hippocampal system that is generally associated with long-term memory retrieval. This is consistent with the finding, mentioned earlier, that memory for the last few list items is spared in Korsakoff’s amnesia ( Baddeley and Warrington, 1970 ; Carlesimo et al., 1995 ). In these studies, the part of the recency effect based on short-term memory could reflect a short amount of time between presentation and recall of the last few items, or it could reflect the absence of interference between presentation and recall of the last few items. Thus, we can say that short-term memory exists, but often without great clarity as to whether the limit is a time limit or a chunk capacity limit.

The short-term memory/working memory distinction

The distinction between short-term memory and working memory is clouded in a bit of confusion but that is largely the result of different investigators using different definitions. Miller et al. (1960) used the term “working memory” to refer to temporary memory from a functional standpoint, so from their point of view there is no clear distinction between short-term and working memory. Baddeley and Hitch (1974) were fairly consistent with this definition but overlaid some descriptions on the terms that distinguished them. They thought of short-term memory as the unitary holding place as described by, for example, Atkinson and Shiffrin (1968) . When they realized that the evidence actually was consistent with a multi-component system that could not be reduced to a unitary short-term store, they used the term working memory to describe that entire system. Cowan (1988) maintained a multi-component view, like Baddeley and Hitch, but without a commitment to precisely their components; instead, the basic subdivisions of working memory were said to be the short-term storage components (activated memory along with the focus of attention within it, shown in Fig. 1 ) and central executive processes that manipulate stored information. By Cowan’s account, Baddeley’s (1986) phonological loop and visuospatial sketchpad would be viewed as just two of many aspects of activated memory, which are susceptible to interference to a degree that depends upon the similarity between features of the activated and interfering information sources. Baddeley’s (2000) episodic buffer is possibly the same as the information saved in Cowan’s focus of attention, or at least is a closely similar concept.

There has been some shift in the definition or description of working memory along with a shift in the explanation of why the newer working memory tasks correlate with intelligence and aptitude measures so much more highly than do simple, traditional, short-term memory tasks such as serial recall. Daneman and Carpenter (1980) had assumed that what is critical is to use working memory tasks that include both storage and processing components, so as to engage all of the parts of working memory as described, for example, by Baddeley and Hitch (1974) . Instead, Engle et al. (1999) and Kane et al. (2001) proposed that what is critical is whether the working memory task is challenging in terms of the control of attention. For example, Kane et al. found that working memory span storage-and-processing tasks correlates well with the ability to inhibit the natural tendency to look toward a suddenly appearing stimulus and instead to look the other way, the antisaccade task. Similarly, Conway et al. (2001) found that individuals scoring high on storage-and-processing tests of working memory notice their names in a channel to be ignored in dichotic listening much less often than low-span individuals; the high-span individuals apparently are better able to make their primary task performance less vulnerable to distraction, but this comes at the expense of being a bit oblivious to irrelevant aspects of their surroundings. In response to such research, Engle and colleagues sometimes used the term working memory to refer only to the processes related to controlling attention. By doing so, their definition of working memory seems at odds with previous definitions but that new definition allows the simple statement that working memory correlates highly with aptitudes, whereas short-term memory (redefined to include only the non-attention-related aspects of memory storage) does not correlate so highly with aptitudes.

Cowan et al. (2006b) , while adhering to the more traditional definition of working memory, made an assertion about working memory similar to that of Engle and colleagues, but a bit more complex. They proposed, on the basis of some developmental and correlational evidence, that multiple functions of attention are relevant to individual differences in aptitudes. The control of attention is relevant, but there is an independent contribution from the number of items that can be held in attention, or its scope. According to this view, what may be necessary for a working memory procedure to correlate well with cognitive aptitudes is that the task must prevent covert verbal rehearsal so that the participant must rely on more attention-demanding processing and/or storage to carry out the task. Cowan et al. (2005) found that the task can be much simpler than the storage-and-processing procedures. For example, in a version of the running memory span test, digits are presented very quickly and the series stops at an unpredictable point, after which the participant is to recall as many items as possible from the end of the list. Rehearsal is impossible and, when the list ends, information presumably must be retrieved from activated sensory or phonological features into the focus of attention. This type of task correlated with aptitudes, as did several other measures of the scope of attention ( Cowan et al., 2005 , 2006b ). In children too young to use covert verbal rehearsal (unlike older children and adults), even a simple digit span task served as an excellent correlate with aptitudes.

Other research verifies this idea that a working memory test will correlate well with cognitive aptitudes to the extent that it requires that attention be used for storage and/or processing. Gavens and Barrouillet (2004) carried out a developmental study in which they controlled the difficulty and duration of a processing task that came between items to be recalled. There still was a developmental difference in span, which they attributed to the development of a basic capacity, which could reflect a developmental increase in the scope of attention (cf. Cowan et al., 2005 ). Lépine et al. (2005) showed that what was important for a storage-and-processing type of span task to correlate well with aptitudes is for the processing component of the task (in this case, reading letters aloud) to occur quickly enough to prevent various types of rehearsal to sneak in between (see also Conlin et al., 2005 ).

Several papers have pitted storage and processing (perhaps the scope versus control of attention?) against one another to see which is more important in accounting for individual differences. Vogel et al. (2005) used a visual array task modified for use with a component of event-related potentials that indicates storage in visual working memory, termed contralateral delay activity (CDA). This activity was found to depend not only on the number of relevant objects in the display (e.g., red bars at varying angles to be remembered), but sometimes also on the number of irrelevant objects to be ignored (e.g., blue bars). For high-span individuals, the CDA for two relevant objects was found to be similar whether or not there also were two irrelevant objects in the display. However, for low-span individuals, the CDA for two relevant objects combined with two irrelevant objects was similar to the CDA for displays with four relevant objects alone, as if the irrelevant objects could not be excluded from working memory. One limitation of the study is that the separation of participants into high versus low span was based on the CDA also, and the task used to measure the CDA inevitably required selective attention (to one half of the display) on every trial, whether or not it included objects of an irrelevant color.

Gold et al. (2006) investigated similar issues in a behavioral design, and testing the difference between schizophrenic patients and normal control participants. Each trial started with a cue to attend to one part of the display at the expense of another (e.g., bars of one relevant color but not another, irrelevant color). The probe display was a set that was cued for relevance on most trials (in some experiments, 75%) whereas, occasionally, the probe display was a set that was not cued. This allowed a separate measure of the control of attention (the advantage for cued items over uncued items) and the storage capacity of working memory (the mean number of items recalled from each array, adding across cued and uncued sets). Unlike the initial expectations, the clear result was that the difference between groups was in the capacity, not in the control of attention. It would be interesting to know whether the same type of result could be obtained for high versus low span normal individuals, or whether that comparison instead would show a control-of-attention difference between these groups as Vogel et al. (2005) must predict. Friedman et al. (2006) found that not all central executive functions correlated with aptitudes; updating working memory did, but inhibition and shifting of attention did not. On the other hand, recall that Cowan et al. (2006b) did find was that a control-of-attention task was related to aptitudes.

In sum, the question of whether short-term memory and working memory are different may be a matter of semantics. There are clearly differences between simple serial recall tasks that do not correlate very well with aptitude tests in adults, and other tasks requiring memory and processing, or memory without the possibility of rehearsal, that correlate much better with aptitudes. Whether to use the term working memory for the latter set of tasks, or whether to reserve that term for the entire system of short-term memory preservation and manipulation, is a matter of taste. The more important, substantive question may be why some tasks correlate with aptitude much better than others.

The distinction between long-term and short-term memory depends on whether it can be demonstrated that there are properties specific to short-term memory; the main candidates include temporal decay and a chunk capacity limit. The question of decay is still pretty much open to debate, whereas there is growing support for a chunk capacity limit. These limits were discussed in a framework shown in Fig. 1 .

The distinction between short-term memory and working memory is one that depends on the definition that one accepts. Nevertheless, the substantive question is why some tests of memory over the short term serve as some of the best correlates of cognitive aptitudes, whereas others do not. The answer seems to point to the importance of an attentional system used both for processing and for storage. The efficiency of this system and its use in working memory seem to differ substantially across individuals (e.g., Conway et al., 2002 ; Kane et al., 2004 ; Cowan et al., 2005 , 2006b ), as well as improving with development in childhood ( Cowan et al., 2005 , 2006b ) and declining in old age ( Naveh-Benjamin et al., 2007 ; Stoltzfus et al., 1996 ; Cowan et al., 2006c ).

Acknowledgment

This work was completed with the assistance of NIH Grant R01 HD-21338.

• Atkinson RC, Shiffrin RM. Human memory: a proposed system and its control processes. In: Spence KW, Spence JT, editors. The Psychology of Learning and Motivation: Advances in Research and Theory. Vol. 2. New York: Academic Press; 1968. pp. 89–195. [ Google Scholar ]
• Baddeley A. The episodic buffer: a new component of working memory? Trends Cogn. Sci. 2000; 4 :417–423. [ PubMed ] [ Google Scholar ]
• Baddeley AD. Oxford Psychology Series No. 11. Oxford: Clarendon Press; 1986. Working Memory. [ Google Scholar ]
• Baddeley AD, Hitch G. Working memory. In: Bower GH, editor. The Psychology of Learning and Motivation. Vol. 8. New York: Academic Press; 1974. pp. 47–89. [ Google Scholar ]
• Baddeley AD, Scott D. Short-term forgetting in the absence of proactive inhibition. Q. J. Exp. Psychol. 1971; 23 :275–283. [ Google Scholar ]
• Baddeley AD, Thomson N, Buchanan M. Word length and the structure of short-term memory. J. Verbal Learn. Verbal Behav. 1975; 14 :575–589. [ Google Scholar ]
• Baddeley AD, Warrington EK. Amnesia and the distinction between long- and short-term memory. J. Verbal Learn. Verbal Behav. 1970; 9 :176–189. [ Google Scholar ]
• Barrouillet P, Bernardin S, Camos V. Time constraints and resource sharing in adults’ working memory spans. J. Exp. Psychol.: Gen. 2004; 133 :83–100. [ PubMed ] [ Google Scholar ]
• Barrouillet P, Bernardin S, Portrat S, Vergauwe E, Camos V. Time and cognitive load in working memory. J. Exp. Psychol. Learn. Mem. Cogn. 2007; 33 :570–585. [ PubMed ] [ Google Scholar ]
• Bjork RA, Whitten WB. Recency-sensitive retrieval processes in long-term free recall. Cogn. Psychol. 1974; 6 :173–189. [ Google Scholar ]
• Broadbent DE. Perception and Communication. New York: Pergamon Press; 1958. [ Google Scholar ]
• Broadbent DE. The magic number seven after fifteen years. In: Kennedy A, Wilkes A, editors. Studies in Long-Term Memory. Oxford, England: Wiley; 1975. pp. 3–18. [ Google Scholar ]
• Brown GDA, Preece T, Hulme C. Oscillator-based memory for serial order. Psychol. Rev. 2000; 107 :127–181. [ PubMed ] [ Google Scholar ]
• Bunting MF, Cowan N, Saults JS. How does running memory span work? Q. J. Exp. Psychol. 2006; 59 :1691–1700. [ PMC free article ] [ PubMed ] [ Google Scholar ]
• Carlesimo GA, Sabbadini M, Fadda L, Caltagirone C. Different components in word-list forgetting of pure amnesics, degenerative demented and healthy subjects. Cortex. 1995; 31 :735–745. [ PubMed ] [ Google Scholar ]
• Chen Z, Cowan N. Chunk limits and length limits in immediate recall: a reconciliation. J. Exp. Psychol. Learn. Mem. Cogn. 2005; 31 :1235–1249. [ PMC free article ] [ PubMed ] [ Google Scholar ]
• Cocchini G, Logie RH, Della Sala S, MacPherson SE, Baddeley AD. Concurrent performance of two memory tasks: evidence for domain-specific working memory systems. Mem. Cogn. 2002; 30 :1086–1095. [ PubMed ] [ Google Scholar ]
• Conlin JA, Gathercole SE, Adams JW. Children’s working memory: investigating performance limitations in complex span tasks. J. Exp. Child Psychol. 2005; 90 :303–317. [ PubMed ] [ Google Scholar ]
• Conway ARA, Cowan N, Bunting MF. The cocktail party phenomenon revisited: the importance of working memory capacity. Psychon. Bull. Rev. 2001; 8 :331–335. [ PubMed ] [ Google Scholar ]
• Conway ARA, Cowan N, Bunting MF, Therriault DJ, Minkoff SRB. A latent variable analysis of working memory capacity, short-term memory capacity, processing speed, and general fluid intelligence. Intelligence. 2002; 30 :163–183. [ Google Scholar ]
• Conway ARA, Kane MJ, Bunting MF, Hambrick DZ, Wilhelm O, Engle RW. Working memory span tasks: a methodological review and user’s guide. Psychon. Bull. Rev. 2005; 12 :769–786. [ PubMed ] [ Google Scholar ]
• Cowan N. Evolving conceptions of memory storage, selective attention, and their mutual constraints within the human information processing system. Psychol. Bull. 1988; 104 :163–191. [ PubMed ] [ Google Scholar ]
• Cowan N. Verbal memory span and the timing of spoken recall. J. Mem. Lang. 1992; 31 :668–684. [ Google Scholar ]
• Cowan N. Oxford Psychology Series No. 26. New York: Oxford University Press; 1995. Attention and memory: an integrated framework. [ Google Scholar ]
• Cowan N. An embedded-processes model of working memory. In: Miyake A, Shah P, editors. Models of Working Memory: Mechanisms of Active Maintenance and Executive Control. Cambridge, UK: Cambridge University Press; 1999. pp. 62–101. [ Google Scholar ]
• Cowan N. The magical number 4 in short-term memory: a reconsideration of mental storage capacity. Behav. Brain Sci. 2001; 24 :87–185. [ PubMed ] [ Google Scholar ]
• Cowan N. Working memory capacity. Hove, East Sussex, UK: Psychology Press; 2005. [ Google Scholar ]
• Cowan N, Aubuchon AM. Psychon. Bull. Rev. Short-term memory loss over time without retroactive stimulus interference. (in press) [ PMC free article ] [ PubMed ] [ Google Scholar ]
• Cowan N, Chen Z, Rouder JN. Constant capacity in an immediate serial-recall task: a logical sequel to Miller (1956) Psychol. Sci. 2004; 15 :634–640. [ PubMed ] [ Google Scholar ]
• Cowan N, Elliott EM, Saults JS, Morey CC, Mattox S, Hismjatullina A, Conway ARA. On the capacity of attention: its estimation and its role in working memory and cognitive aptitudes. Cogn. Psychol. 2005; 51 :42–100. [ PMC free article ] [ PubMed ] [ Google Scholar ]
• Cowan N, Elliott EM, Saults JS, Nugent LD, Bomb P, Hismjatullina A. Rethinking speed theories of cognitive development: increasing the rate of recall without affecting accuracy. Psychol. Sci. 2006a; 17 :67–73. [ PMC free article ] [ PubMed ] [ Google Scholar ]
• Cowan N, Fristoe NM, Elliott EM, Brunner RP, Saults JS. Scope of attention, control of attention, and intelligence in children and adults. Mem. Cogn. 2006b; 34 :1754–1768. [ PMC free article ] [ PubMed ] [ Google Scholar ]
• Cowan N, Morey CC. How can dual-task working memory retention limits be investigated? Psychol. Sci. 2007; 18 :686–688. [ PMC free article ] [ PubMed ] [ Google Scholar ]
• Cowan N, Naveh-Benjamin M, Kilb A, Saults JS. Life-span development of visual working memory: when is feature binding difficult? Dev. Psychol. 2006c; 42 :1089–1102. [ PMC free article ] [ PubMed ] [ Google Scholar ]
• Cowan N, Nugent LD, Elliott EM, Ponomarev I, Saults JS. The role of attention in the development of short-term memory: age differences in the verbal span of apprehension. Child Dev. 1999; 70 :1082–1097. [ PubMed ] [ Google Scholar ]
• Cowan N, Saults JS, Nugent LD. The role of absolute and relative amounts of time in forgetting within immediate memory: the case of tone pitch comparisons. Psychon. Bull. Rev. 1997a; 4 :393–397. [ Google Scholar ]
• Cowan N, Wood NL, Nugent LD, Treisman M. There are two word length effects in verbal short-term memory: opposed effects of duration and complexity. Psychol. Sci. 1997b; 8 :290–295. [ Google Scholar ]
• Craik FIM, Birtwistle J. Proactive inhibition in free recall. J. Exp. Psychol. 1971; 91 :120–123. [ Google Scholar ]
• Crowder RG. The demise of short-term memory. Acta Psychol. 1982; 50 :291–323. [ PubMed ] [ Google Scholar ]
• Crowder RG. Short-term memory: where do we stand? Mem. Cogn. 1993; 21 :142–145. [ PubMed ] [ Google Scholar ]
• Daneman M, Carpenter PA. Individual differences in working memory and reading. J Verbal Learn. Verbal Behav. 1980; 19 :450–466. [ Google Scholar ]
• Daneman M, Merikle PM. Working memory and language comprehension: a meta-analysis. Psychon. Bull. Rev. 1996; 3 :422–433. [ PubMed ] [ Google Scholar ]
• Darwin CJ, Turvey MT, Crowder RG. An auditory analogue of the Sperling partial report procedure: evidence for brief auditory storage. Cogn. Psychol. 1972; 3 :255–267. [ Google Scholar ]
• Davelaar EJ, Goshen-Gottstein Y, Ashkenazi A, Haarman HJ, Usher M. The demise of short-term memory revisited: empirical and computational investigations of recency effects. Psychol. Rev. 2005; 112 :3–42. [ PubMed ] [ Google Scholar ]
• Ebbinghaus H. Translated by H.A. Ruger and C.E. Bussenius. New York: Teachers College, Columbia University; 18851913. Memory: a contribution to experimental psychology. (Originally in German, Ueber das gedächtnis: Untersuchen zur experimentellen psychologie) [ Google Scholar ]
• Engle RW. Working memory capacity as executive attention. Curr. Dir. Psychol. Sci. 2002; 11 :19–23. [ Google Scholar ]
• Engle RW, Tuholski SW, Laughlin JE, Conway ARA. Working memory, short-term memory, and general fluid intelligence: a latent-variable approach. J. Exp. Psychol. Gen. 1999; 128 :309–331. [ PubMed ] [ Google Scholar ]
• Ericsson KA, Kintsch W. Long-term working memory. Psychol. Rev. 1995; 102 :211–245. [ PubMed ] [ Google Scholar ]
• Friedman NP, Miyake A, Corley RP, Young SE, DeFries JC, Hewitt JK. Not all executive functions are related to intelligence. Psychol. Sci. 2006; 17 :172–179. [ PubMed ] [ Google Scholar ]
• Gavens N, Barrouillet P. Delays of retention, processing efficiency, and attentional resources in working memory span development. J. Mem. Lang. 2004; 51 :644–657. [ Google Scholar ]
• Glanzer M, Cunitz AR. Two storage mechanisms in free recall. J. Verbal Learn. Verbal Behav. 1966; 5 :351–360. [ Google Scholar ]
• Glenberg AM, Swanson NC. A temporal distinctiveness theory of recency and modality effects. J. Exp. Psychol. Learn. Mem. Cogn. 1986; 12 :3–15. [ PubMed ] [ Google Scholar ]
• Gold JM, Fuller RL, Robinson BM, McMahon RP, Braun EL, Luck SJ. Intact attentional control of working memory encoding in schizophrenia. J. Abnorm. Psychol. 2006; 115 :658–673. [ PubMed ] [ Google Scholar ]
• Greene RL. The influence of experimental design: the example of the Brown-Peterson paradigm. Can. J. Exp. Psychol. 1996; 50 :240–242. [ Google Scholar ]
• Guttentag RE. The mental effort requirement of cumulative rehearsal: a developmental study. J. Exp. Child Psychol. 1984; 37 :92–106. [ Google Scholar ]
• Hebb DO. Organization of Behavior. New York: Wiley; 1949. [ Google Scholar ]
• Hockey R. Rate of presentation in running memory and direct manipulation of input-processing strategies. Q. J. Exp. Psychol. A. 1973; 25 :104–111. [ Google Scholar ]
• James W. The Principles of Psychology. New York: Henry Holt; 1890. [ Google Scholar ]
• Kane MJ, Bleckley MK, Conway ARA, Engle RW. A controlled-attention view of working-memory capacity. J. Exp. Psychol. Gen. 2001; 130 :169–183. [ PubMed ] [ Google Scholar ]
• Kane MJ, Hambrick DZ, Tuholski SW, Wilhelm O, Payne TW, Engle RE. The generality of working-memory capacity: a latent-variable approach to verbal and visuo-spatial memory span and reasoning. J. Exp. Psychol. Gen. 2004; 133 :189–217. [ PubMed ] [ Google Scholar ]
• Keppel G, Underwood BJ. Proactive inhibition in short-term retention of single items. J. Verbal Learn. Verbal Behav. 1962; 1 :153–161. [ Google Scholar ]
• Kyllonen PC, Christal RE. Reasoning ability is (little more than) working-memory capacity? Intelligence. 1990; 14 :389–433. [ Google Scholar ]
• Lépine R, Barrouillet P, Camos V. What makes working memory spans so predictive of high level cognition? Psychon. Bull. Rev. 2005; 12 :165–170. [ PubMed ] [ Google Scholar ]
• Lewandowsky S, Duncan M, Brown GDA. Time does not cause forgetting in short-term serial recall. Psychon. Bull. Rev. 2004; 11 :771–790. [ PubMed ] [ Google Scholar ]
• McGeoch JA. Forgetting and the law of disuse. Psychol. Rev. 1932; 39 :352–370. [ Google Scholar ]
• Melton AW. Implications of short-term memory for a general theory of memory. J. Verbal Learn. Verbal Behav. 1963; 2 :1–21. [ Google Scholar ]
• Miller GA. The magical number seven, plus or minus two: some limits on our capacity for processing information. Psychol. Rev. 1956; 63 :81–97. [ PubMed ] [ Google Scholar ]
• Miller GA. George A. Miller. In: Lindzey G, editor. A History of Psychology in Autobiography. Vol. VIII. Stanford, CA: Stanford University Press; 1989. pp. 391–418. [ Google Scholar ]
• Miller GA, Galanter E, Pribram KH. Plans and the structure of behavior. New York: Holt, Rinehart and Winston, Inc; 1960. [ Google Scholar ]
• Morey CC, Cowan N. When visual and verbal memories compete: evidence of cross-domain limits in working memory. Psychon. Bull. Rev. 2004; 11 :296–301. [ PubMed ] [ Google Scholar ]
• Morey CC, Cowan N. When do visual and verbal memories conflict? The importance of working-memory load and retrieval. J. Exp. Psychol. Learn. Mem. Cogn. 2005; 31 :703–713. [ PMC free article ] [ PubMed ] [ Google Scholar ]
• Nairne JS. Remembering over the short-term: the case against the standard model. Annu. Rev. Psychol. 2002; 53 :53–81. [ PubMed ] [ Google Scholar ]
• Naveh-Benjamin M, Cowan N, Kilb A, Chen Z. Age-related differences in immediate serial recall: dissociating chunk formation and capacity. Mem. Cognit. 2007; 35 :724–737. [ PMC free article ] [ PubMed ] [ Google Scholar ]
• Neath I, Surprenant A. Human memory. 2nd ed. Belmont, CA: Wadsworth; 2003. [ Google Scholar ]
• Nipher FE. On the distribution of errors in numbers written from memory. Trans. Acad. Sci. St. Louis. 1878; 3 :ccx–ccxi. [ Google Scholar ]
• Peterson LR, Peterson MJ. Short-term retention of individual verbal items. J. Exp. Psychol. 1959; 58 :193–198. [ PubMed ] [ Google Scholar ]
• Saults JS, Cowan N. A central capacity limit to the simultaneous storage of visual and auditory arrays in working memory. J. Exp. Psychol. 2007; 136 :663–684. [ PMC free article ] [ PubMed ] [ Google Scholar ]
• Sperling G. The information available in brief visual presentations. Psychol. Monogr. 1960; 74 (Whole No. 498.) [ Google Scholar ]
• Stoltzfus ER, Hasher L, Zacks RT. Working memory and retrieval: an inhibition-resource approach. In: Richardson JTE, Engle RW, Hasher L, Logie RH, Stoltzfus ER, Zacks RT, editors. Working Memory and Human Cognition. New York: Oxford University Press; 1996. pp. 66–88. [ Google Scholar ]
• Talmi D, Grady CL, Goshen-Gottstein Y, Moscovitch M. Neuroimaging the serial position curve: a test of single-store versus dual-store models. Psychol. Sci. 2005; 16 :716–723. [ PubMed ] [ Google Scholar ]
• Tulving E, Patkau JE. Concurrent effects of contextual constraint and word frequency on immediate recall and learning of verbal material. Can. J. Psychol. 1962; 16 :83–95. [ PubMed ] [ Google Scholar ]
• Turvey MT, Brick P, Osborn J. Proactive interference in short-term memory as a function of prior-item retention interval. Q. J. Exp. Psychol. 1970; 22 :142–147. [ Google Scholar ]
• Tzeng OJL. Positive recency effect in a delayed free recall. J. Verbal Learn. Verbal Behav. 1973; 12 :436–439. [ Google Scholar ]
• Unsworth N, Engle RW. The nature of individual differences in working memory capacity: active maintenance in primary memory and controlled search from secondary memory. Psychol. Rev. 2007; 114 :104–132. [ PubMed ] [ Google Scholar ]
• Vogel EK, McCollough AW, Machizawa MG. Neural measures reveal individual differences in controlling access to working memory. Nature. 2005; 438 :500–503. [ PubMed ] [ Google Scholar ]
• Vogel EK, Woodman GF, Luck SJ. The time course of consolidation in visual working memory. J. Exp. Psychol. Hum. Percept. Perform. 2006; 32 :1436–1451. [ PubMed ] [ Google Scholar ]
• Waugh NC, Norman DA. Primary memory. Psychol. Rev. 1965; 72 :89–104. [ PubMed ] [ Google Scholar ]
• Wickelgren WA. Single-trace fragility theory of memory dynamics. Mem. Cogn. 1974; 2 :775–780. [ PubMed ] [ Google Scholar ]

Help | Advanced Search

Computer Science > Machine Learning

Title: long short-term memory rnn.

Abstract: This paper is based on a machine learning project at the Norwegian University of Science and Technology, fall 2020. The project was initiated with a literature review on the latest developments within time-series forecasting methods in the scientific community over the past five years. The paper summarizes the essential aspects of this research. Furthermore, in this paper, we introduce an LSTM cell's architecture, and explain how different components go together to alter the cell's memory and predict the output. Also, the paper provides the necessary formulas and foundations to calculate a forward iteration through an LSTM. Then, the paper refers to some practical applications and research that emphasize the strength and weaknesses of LSTMs, shown within the time-series domain and the natural language processing (NLP) domain. Finally, alternative statistical methods for time series predictions are highlighted, where the paper outline ARIMA and exponential smoothing. Nevertheless, as LSTMs can be viewed as a complex architecture, the paper assumes that the reader has some knowledge of essential machine learning aspects, such as the multi-layer perceptron, activation functions, overfitting, backpropagation, bias, over- and underfitting, and more.

Submission history

Access paper:.

• Other Formats

References & Citations

• Google Scholar
• Semantic Scholar

DBLP - CS Bibliography

Bibtex formatted citation.

Bibliographic and Citation Tools

Code, data and media associated with this article, recommenders and search tools.

• Institution

arXivLabs: experimental projects with community collaborators

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs .

Information

• Author Services

Initiatives

You are accessing a machine-readable page. In order to be human-readable, please install an RSS reader.

All articles published by MDPI are made immediately available worldwide under an open access license. No special permission is required to reuse all or part of the article published by MDPI, including figures and tables. For articles published under an open access Creative Common CC BY license, any part of the article may be reused without permission provided that the original article is clearly cited. For more information, please refer to https://www.mdpi.com/openaccess .

Feature papers represent the most advanced research with significant potential for high impact in the field. A Feature Paper should be a substantial original Article that involves several techniques or approaches, provides an outlook for future research directions and describes possible research applications.

Feature papers are submitted upon individual invitation or recommendation by the scientific editors and must receive positive feedback from the reviewers.

Editor’s Choice articles are based on recommendations by the scientific editors of MDPI journals from around the world. Editors select a small number of articles recently published in the journal that they believe will be particularly interesting to readers, or important in the respective research area. The aim is to provide a snapshot of some of the most exciting work published in the various research areas of the journal.

Original Submission Date Received: .

• Active Journals
• Find a Journal
• Proceedings Series
• For Authors
• For Reviewers
• For Editors
• For Librarians
• For Publishers
• For Societies
• For Conference Organizers
• Open Access Policy
• Institutional Open Access Program
• Special Issues Guidelines
• Editorial Process
• Research and Publication Ethics
• Article Processing Charges
• Testimonials
• Preprints.org
• SciProfiles
• Encyclopedia

Article Menu

• Subscribe SciFeed
• Recommended Articles
• Google Scholar
• on Google Scholar
• Table of Contents

Find support for a specific problem in the support section of our website.

Please let us know what you think of our products and services.

Visit our dedicated information section to learn more about MDPI.

JSmol Viewer

Using machine learning models for short-term prediction of dissolved oxygen in a microtidal estuary.

1. Introduction

2. materials and methods, 2.1. study area and data, 2.2. machine learning models, 2.2.1. the multi-layer perceptron (mlp), 2.2.2. the recurrent neural network (rnn), 2.2.3. long short-term memory (lstm) networks, 2.2.4. gradient boosting (gb), 2.2.5. autokeras, 2.3. the model application process, 2.4. model parameter tuning, 4. discussion, 5. conclusions, author contributions, data availability statement, acknowledgments, conflicts of interest.

• Paerl, H.W.; Rossignol, K.L.; Hall, S.N.; Peierls, B.L.; Wetz, M.S. Phytoplankton Community Indicators of Short- and Long-Term Ecological Change in the Anthropogenically and Climatically Impacted Neuse River Estuary, North Carolina, USA. Estuaries Coasts 2010 , 33 , 485–497. [ Google Scholar ] [ CrossRef ]
• Latif, S.D.; Azmi, M.S.B.N.; Ahmed, A.N.; Fai, C.M.; El-Shafie, A. Application of Artificial Neural Network for Forecasting Nitrate Concentration as a Water Quality Parameter: A Case Study of Feitsui Reservoir, Taiwan. IJDNE 2020 , 15 , 647–652. [ Google Scholar ] [ CrossRef ]
• Ziyad Sami, B.F.; Latif, S.D.; Ahmed, A.N.; Chow, M.F.; Murti, M.A.; Suhendi, A.; Ziyad Sami, B.H.; Wong, J.K.; Birima, A.H.; El-Shafie, A. Machine Learning Algorithm as a Sustainable Tool for Dissolved Oxygen Prediction: A Case Study of Feitsui Reservoir, Taiwan. Sci. Rep. 2022 , 12 , 3649. [ Google Scholar ] [ CrossRef ]
• Zhi, W.; Feng, D.; Tsai, W.-P.; Sterle, G.; Harpold, A.; Shen, C.; Li, L. From Hydrometeorology to River Water Quality: Can a Deep Learning Model Predict Dissolved Oxygen at the Continental Scale? Environ. Sci. Technol. 2021 , 55 , 2357–2368. [ Google Scholar ] [ CrossRef ]
• Vaquer-Sunyer, R.; Duarte, C.M. Thresholds of Hypoxia for Marine Biodiversity. Proc. Natl. Acad. Sci. USA 2008 , 105 , 15452–15457. [ Google Scholar ] [ CrossRef ]
• Farrell, A.P.; Richards, J.G. Chapter 11 Defining Hypoxia: An Integrative Synthesis of the Responses of Fish to Hypoxia. In Fish Physiology ; Richards, J.G., Farrell, A.P., Brauner, C.J., Eds.; Hypoxia; Academic Press: Cambridge, MA, USA, 2009; Volume 27, pp. 487–503. [ Google Scholar ]
• Biddanda, B.A.; Weinke, A.D.; Kendall, S.T.; Gereaux, L.C.; Holcomb, T.M.; Snider, M.J.; Dila, D.K.; Long, S.A.; VandenBerg, C.; Knapp, K.; et al. Chronicles of Hypoxia: Time-Series Buoy Observations Reveal Annually Recurring Seasonal Basin-Wide Hypoxia in Muskegon Lake—A Great Lakes Estuary. J. Great Lakes Res. 2018 , 44 , 219–229. [ Google Scholar ] [ CrossRef ]
• Rowe, M.D.; Anderson, E.J.; Wynne, T.T.; Stumpf, R.P.; Fanslow, D.L.; Kijanka, K.; Vanderploeg, H.A.; Strickler, J.R.; Davis, T.W. Vertical Distribution of Buoyant Microcystis Blooms in a Lagrangian Particle Tracking Model for Short-Term Forecasts in Lake Erie. J. Geophys. Res. Ocean. 2016 , 175 , 238. [ Google Scholar ] [ CrossRef ]
• Moshogianis, A. A Statistical Model for the Prediction of Dissolved Oxygen Dynamics and the Potential for Hypoxia in the Mississippi Sound and Bight. Master’s Thesis, University of Southern Mississippi, Hattiesburg, MS, USA, 2015. [ Google Scholar ]
• Katin, A.; Del Giudice, D.; Obenour, D.R. Temporally Resolved Coastal Hypoxia Forecasting and Uncertainty Assessment via Bayesian Mechanistic Modeling. Hydrol. Earth Syst. Sci. 2022 , 26 , 1131–1143. [ Google Scholar ] [ CrossRef ]
• Chubarenko, I.; Tchepikova, I. Modelling of Man-Made Contribution to Salinity Increase into the Vistula Lagoon (Baltic Sea). Ecol. Model. 2001 , 138 , 87–100. [ Google Scholar ] [ CrossRef ]
• Marcomini, A.; Sute, G.W., II; Critto, A. (Eds.) Decision Support Systems for Risk-Based Management of Contaminated Sites ; Springer Science & Business Media: Berlin/Heidelberg, Germany, 2008; ISBN 978-0-387-09722-0. [ Google Scholar ]
• Scavia, D.; Justić, D.; Obenour, D.R.; Craig, J.K.; Wang, L. Hypoxic Volume Is More Responsive than Hypoxic Area to Nutrient Load Reductions in the Northern Gulf of Mexico—And It Matters to Fish and Fisheries. Environ. Res. Lett. 2019 , 14 , 024012. [ Google Scholar ] [ CrossRef ]
• Borsuk, M.E.; Higdon, D.; Stow, C.A.; Reckhow, K.H. A Bayesian Hierarchical Model to Predict Benthic Oxygen Demand from Organic Matter Loading in Estuaries and Coastal Zones. Ecol. Model. 2001 , 143 , 165–181. [ Google Scholar ] [ CrossRef ]
• Katin, A.; Del Giudice, D.; Obenour, D.R. Modeling Biophysical Controls on Hypoxia in a Shallow Estuary Using a Bayesian Mechanistic Approach. Environ. Model. Softw. 2019 , 120 , 104491. [ Google Scholar ] [ CrossRef ]
• Ahmed, A.A.M. Prediction of Dissolved Oxygen in Surma River by Biochemical Oxygen Demand and Chemical Oxygen Demand Using the Artificial Neural Networks (ANNs). J. King Saud. Univ.—Eng. Sci. 2017 , 29 , 151–158. [ Google Scholar ] [ CrossRef ]
• Yu, X.; Shen, J.; Du, J. A Machine-Learning-Based Model for Water Quality in Coastal Waters, Taking Dissolved Oxygen and Hypoxia in Chesapeake Bay as an Example. Water Resour. Res. 2020 , 56 , e2020WR027227. [ Google Scholar ] [ CrossRef ]
• Agatonovic-Kustrin, S.; Beresford, R. Basic Concepts of Artificial Neural Network (ANN) Modeling and Its Application in Pharmaceutical Research. J. Pharm. Biomed. Anal. 2000 , 22 , 717–727. [ Google Scholar ] [ CrossRef ]
• Montgomery, D.C.; Peck, E.A.; Vining, G.G. Introduction to Linear Regression Analysis , 6th ed.; Wiley Series in Probability and Statistics; Wiley: Hoboken, NJ, USA, 2021; ISBN 978-1-119-57872-7. [ Google Scholar ]
• Steinwart, I.; Christmann, A. Support Vector Machines ; Springer: Berlin/Heidelberg, Germany, 2008. [ Google Scholar ]
• Lu, H.; Ma, X. Hybrid Decision Tree-Based Machine Learning Models for Short-Term Water Quality Prediction. Chemosphere 2020 , 249 , 126169. [ Google Scholar ] [ CrossRef ]
• Asadollah, S.B.H.S.; Sharafati, A.; Motta, D.; Yaseen, Z.M. River Water Quality Index Prediction and Uncertainty Analysis: A Comparative Study of Machine Learning Models. J. Environ. Chem. Eng. 2021 , 9 , 104599. [ Google Scholar ] [ CrossRef ]
• Maier, H.R.; Jain, A.; Dandy, G.C.; Sudheer, K.P. Methods Used for the Development of Neural Networks for the Prediction of Water Resource Variables in River Systems: Current Status and Future Directions. Environ. Model. Softw. 2010 , 25 , 891–909. [ Google Scholar ] [ CrossRef ]
• Antanasijević, D.; Pocajt, V.; Povrenović, D.; Perić-Grujić, A.; Ristić, M. Modelling of Dissolved Oxygen Content Using Artificial Neural Networks: Danube River, North Serbia, Case Study: Environmental Science & Pollution Research. Environ. Sci. Pollut. Res. 2013 , 20 , 9006–9013. [ Google Scholar ] [ CrossRef ]
• Huang, J.; Liu, S.; Hassan, S.G.; Xu, L.; Huang, C. A Hybrid Model for Short-Term Dissolved Oxygen Content Prediction. Comput. Electron. Agric. 2021 , 186 , 106216. [ Google Scholar ] [ CrossRef ]
• Nair, J.P. Analysing and Modelling Dissolved Oxygen Concentration Using Deep Learning Architectures. Int. J. Mech. Eng. 2022 , 7 , 12–22. [ Google Scholar ]
• Bentéjac, C.; Csörgő, A.; Martínez-Muñoz, G. A Comparative Analysis of XGBoost. Artif. Intell. Rev. 2021 , 54 , 1937–1967. [ Google Scholar ] [ CrossRef ]
• Sherstinsky, A. Fundamentals of Recurrent Neural Network (RNN) and Long Short-Term Memory (LSTM) Network. Phys. D Nonlinear Phenom. 2020 , 404 , 132306. [ Google Scholar ] [ CrossRef ]
• Taud, H.; Mas, J.F. Multilayer Perceptron (MLP) ; Springer: Berlin/Heidelberg, Germany, 2017; Volume 2024. [ Google Scholar ]
• Jin, H.; Song, Q.; Hu, X. Auto-Keras: An Efficient Neural Architecture Search System. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, Anchorage, AK, USA, 4–8 August 2019; Association for Computing Machinery: New York, NY, USA, 2019; pp. 1946–1956. [ Google Scholar ]
• Thompson, P.A.; Paerl, H.W.; Campbell, L.; Yin, K.; McDonald, K.S. Tropical Cyclones: What Are Their Impacts on Phytoplankton Ecology? J. Plankton Res. 2023 , 45 , 180–204. [ Google Scholar ] [ CrossRef ]
• Stow, C.A.; Roessler, C.; Borsuk, M.E.; Bowen, J.D.; Reckhow, K.H. Comparison of Estuarine Water Quality Models for Total Maximum Daily Load Development in Neuse River Estuary. J. Water Resour. Plann. Manag. 2003 , 129 , 307–314. [ Google Scholar ] [ CrossRef ]
• Paerl, H.; Pinckney, J.; Fear, J.; Peierls, B. Ecosystem Responses to Internal and Watershed Organic Matter Loading:Consequences for Hypoxia in the Eutrophying Neuse River Estuary, North Carolina, USA. Mar. Ecol. Prog. Ser. 1998 , 166 , 17–25. [ Google Scholar ] [ CrossRef ]
• Wool, T.A.; Davie, S.R.; Rodriguez, H.N. Development of Three-Dimensional Hydrodynamic and Water Quality Models to Support Total Maximum Daily Load Decision Process for the Neuse River Estuary, North Carolina. J. Water Resour. Plann. Manage. 2003 , 129 , 295–306. [ Google Scholar ] [ CrossRef ]
• Lin, J.; Liu, Q.; Song, Y.; Liu, J.; Yin, Y.; Hall, N.S. Temporal Prediction of Coastal Water Quality Based on Environmental Factors with Machine Learning. J. Mar. Sci. Eng. 2023 , 11 , 1608. [ Google Scholar ] [ CrossRef ]
• Peirce, C.S. The numerical measure of the success of predictions. Science 1884 , 4 , 453–454. [ Google Scholar ] [ CrossRef ]
• Raheli, B.; Aalami, M.; El-Shafie, A.; Ghorbani, M.; Deo, R. Uncertainty Assessment of the Multilayer Perceptron (MLP) Neural Network Model with Implementation of the Novel Hybrid MLP-FFA Method for Prediction of Biochemical Oxygen Demand and Dissolved Oxygen: A Case Study of Langat River. Environ. Earth Sci. 2017 , 76 , 503. [ Google Scholar ] [ CrossRef ]
• Ismail, M.R.; Awang, M.K.; Rahman, M.N.A.; Makhtar, M. A Multi-Layer Perceptron Approach for Customer Churn Prediction. Int. J. Multimed. Ubiquitous Eng. 2015 , 10 , 213–222. [ Google Scholar ] [ CrossRef ]
• Niroobakhsh, M. Prediction of Water Quality Parameter in Jajrood River Basin: Application of Multi Layer Perceptron (MLP) Perceptron and Radial Basis Function Networks of Artificial Neural Networks (ANNs). Afr. J. Agric. Res. 2012 , 7 , 4131–4139. [ Google Scholar ] [ CrossRef ]
• Selvin, S.; Ravi, V.; Gopalakrishnan, E.A.; Menon, V.; Kp, S. Stock. In Price Prediction Using. LSTM, RNN and CNN-Sliding Window Model. In Proceedings of the 2017 International Conference on Advances in Computing, Communications and Informatics (ICACCI), Udupi, India, 13–16 September 2017; p. 1647. [ Google Scholar ]
• Che, Z.; Purushotham, S.; Cho, K.; Sontag, D.; Liu, Y. Recurrent Neural Networks for Multivariate Time Series with Missing Values. Sci. Rep. 2018 , 8 , 6085. [ Google Scholar ] [ CrossRef ]
• Hu, Z.; Zhang, Y.; Zhao, Y.; Xie, M.; Zhong, J.; Tu, Z.; Liu, J. A Water Quality Prediction Method Based on the Deep LSTM Network Considering Correlation in Smart Mariculture. Sensors 2019 , 19 , 1420. [ Google Scholar ] [ CrossRef ]
• Friedman, J.H. Stochastic Gradient Boosting. Comput. Stat. Data Anal. 2002 , 38 , 367–378. [ Google Scholar ] [ CrossRef ]
• Bolick, M.M.; Post, C.J.; Naser, M.-Z.; Mikhailova, E.A. Comparison of Machine Learning Algorithms to Predict Dissolved Oxygen in an Urban Stream. Env. Sci. Pollut. Res. 2023 , 30 , 78075–78096. [ Google Scholar ] [ CrossRef ]
• Prasad, D.V.V.; Venkataramana, L.Y.; Kumar, P.S.; Prasannamedha, G.; Harshana, S.; Srividya, S.J.; Harrinei, K.; Indraganti, S. Analysis and Prediction of Water Quality Using Deep Learning and Auto Deep Learning Techniques. Sci. Total Environ. 2022 , 821 , 153311. [ Google Scholar ] [ CrossRef ]

Click here to enlarge figure

Share and Cite

Gachloo, M.; Liu, Q.; Song, Y.; Wang, G.; Zhang, S.; Hall, N. Using Machine Learning Models for Short-Term Prediction of Dissolved Oxygen in a Microtidal Estuary. Water 2024 , 16 , 1998. https://doi.org/10.3390/w16141998

Gachloo M, Liu Q, Song Y, Wang G, Zhang S, Hall N. Using Machine Learning Models for Short-Term Prediction of Dissolved Oxygen in a Microtidal Estuary. Water . 2024; 16(14):1998. https://doi.org/10.3390/w16141998

Gachloo, Mina, Qianqian Liu, Yang Song, Guozhi Wang, Shuhao Zhang, and Nathan Hall. 2024. "Using Machine Learning Models for Short-Term Prediction of Dissolved Oxygen in a Microtidal Estuary" Water 16, no. 14: 1998. https://doi.org/10.3390/w16141998

Article Metrics

Article access statistics, further information, mdpi initiatives, follow mdpi.

Subscribe to receive issue release notifications and newsletters from MDPI journals