Working Memory

Abstract

Working memory is the limited-capacity system that holds and manipulates information over intervals of seconds, the workspace in which the contents of thought are held while they are transformed. This article traces the construct from Baddeley and Hitch's multicomponent model, with its phonological loop, visuospatial sketchpad, central executive, and episodic buffer, through the long argument over how much it holds, from Miller's seven to Cowan's four, to the modern study of visual working memory and the unresolved dispute between discrete slots and a continuous resource. It closes on the neural evidence, from the persistent prefrontal firing that Fuster and Goldman-Rakic recorded to the activity-silent codes proposed more recently. Three interactive demonstrations let the reader drive the word-length effect, estimate visual capacity from change detection with Cowan's formula, and compare the precision that slot and resource models predict.

Keywords: working memory, capacity, central executive, visual working memory, prefrontal cortex

Working memory is the system that maintains a limited amount of information in an immediately accessible state while it is used to guide behaviour, distinguished from passive short-term storage by the manipulation it performs on what it holds (Baddeley, 2003). In the Medical Subject Headings vocabulary it has no descriptor of its own: the term is indexed as an entry term under Memory, Short-Term (descriptor D008570), a cataloguing choice that predates the sharp modern distinction between the two and that this article treats as a historical artefact rather than a claim about identity. The construct sits at the centre of cognitive psychology because so much else runs through it: reasoning, language comprehension, and problem solving all require that intermediate products be held while further operations are applied to them, and individual differences in working-memory capacity predict performance across that entire range (Engle, 2002). The sections below set out the dominant architectural model, the century-long argument over capacity, the special case of visual working memory that has driven the field's recent theory, and the neural machinery that appears to implement maintenance.

What Working Memory Is

The defining feature of working memory is in its name: it is memory put to work. A store that merely retained a telephone number until it could be dialled would be short-term memory; a system that retained the number while reversing it, or while using it to compute a total, is working memory, because it operates on its contents rather than only preserving them (Baddeley, 2003). This distinction is not a terminological nicety. The standard laboratory contrast makes it concrete: repeating a span of digits in the order heard is a short-term storage task, whereas repeating them in reverse order is a working-memory task, because the digits must be held and transformed at the same time. The two dissociate in patients and in the factor structure of individual differences, and the tasks that best predict reasoning are the manipulation tasks, not the storage ones (Daneman & Carpenter, 1980).

An influential strand of theory locates the essence of working memory not in storage at all but in the control of attention. On Engle's account, working-memory capacity is the ability to maintain task-relevant information in an active, accessible state in the face of interference and distraction, and the storage limit is a downstream consequence of that attentional limit rather than its cause (Engle, 2002). This is why working-memory span predicts performance on tasks with no obvious memory component, from reading comprehension to the control of eye movements: what the high-capacity individual has more of is not slots but the executive control that keeps the relevant thing in mind. Table 1 sets out the components of the model that has organised most of this work.

Table 1

Components of the Multicomponent Model of Working Memory

Note. The first three components are the original 1974 model; the episodic buffer was added in 2000 to explain how information is integrated across the subsystems and with long-term memory (Baddeley, 2000).

The Multicomponent Model

The modern concept of working memory begins with a deliberate break from the single short-term store of earlier models. Alan Baddeley and Graham Hitch reasoned that if short-term memory were one undifferentiated buffer, loading it with a string of digits should cripple any concurrent task that depended on it; yet participants asked to hold six digits could still reason, comprehend, and learn with only modest cost (Baddeley & Hitch, 1974). They concluded that the system is not a single store but a set of components: a controlling central executive served by two subsidiary systems, one for speech-based and one for visuospatial material. The architecture is drawn in Figure 1.

The best-characterised component is the phonological loop, and its evidence is unusually clean. Two effects reveal its inner structure. The phonological-similarity effect, discovered by Conrad, is that lists of similar-sounding letters (B, C, D, T) are harder to recall in order than dissimilar ones (F, K, Q, R), even when the letters are presented visually, showing that the store codes material by sound rather than by shape (Conrad, 1964). The word-length effect is that immediate memory span is smaller for long words than for short ones, and Baddeley, Thomson, and Buchanan showed that span corresponds closely to the number of items a person can articulate in about two seconds, as if the loop were a tape of fixed duration continually refreshed by rehearsal (Baddeley, Thomson, & Buchanan, 1975). The demonstration below drives this relationship directly. Far from being a laboratory curiosity, the loop appears to be an evolved device for language acquisition: its capacity to hold unfamiliar sound sequences predicts the ability to learn new vocabulary, in children and in second-language learners alike (Baddeley, Gathercole, & Papagno, 1998).

The model's one major revision came in 2000. The original three components could not explain how information was combined, how the sound of a word met its meaning, or how a sentence that exceeded the loop's span was still recalled better than a list of unrelated words. Baddeley added a fourth component, the episodic buffer, a limited store that binds information from the phonological loop, the sketchpad, and long-term memory into integrated, multidimensional episodes under the control of the executive (Baddeley, 2000). The buffer is the model's account of the moment-to-moment unity of conscious experience, and it remains the least specified of the four components (Baddeley, 2003).

Capacity

No question about working memory has been asked more often, or answered more variously, than how much it holds. The famous first answer was George Miller's: across absolute judgement, immediate memory, and other tasks, performance broke down at around seven items, the magical number seven, plus or minus two (Miller, 1956). Miller himself was cautious, half-suspecting the recurrence of seven to be a coincidence, and his deeper contribution was the concept of the chunk: the unit of capacity is not the item but the meaningful group, so that recoding a string of digits into a few familiar clusters expands what can be held without violating the limit. That insight is also what undermines the number seven, because if capacity is counted in chunks, and participants spontaneously chunk, then an uncontrolled span measures chunking ability as much as raw capacity.

When chunking is controlled, the estimate falls. Nelson Cowan surveyed a wide range of procedures designed to prevent recoding, from rapid presentation to tasks that block rehearsal, and concluded that the pure capacity of working memory is about four items, not seven, a limit he argued reflects the number of items that can be held simultaneously in the focus of attention (Cowan, 2001). This magical number four has proved robust, and it recurs beyond verbal memory in visual storage and in the parallel individuation that also caps subitizing and object tracking at roughly the same value.

Measuring capacity well enough to compare individuals demanded better tasks than simple span. The complex span was the breakthrough: Meredyth Daneman and Patricia Carpenter interleaved storage with processing, asking readers to comprehend a series of sentences while retaining the last word of each, and found that this reading span, unlike simple digit span, predicted reading comprehension strongly (Daneman & Carpenter, 1980). The complex-span family that grew from it became the standard instrument of the individual-differences literature, and a methodological review established how the tasks should be scored and how reliable they are (Conway et al., 2005). The upshot is that working-memory capacity, properly measured, is one of the most predictive constructs in cognitive psychology, correlated with reasoning, comprehension, and executive control (Engle, 2002).

Visual Working Memory

The study of capacity was transformed when the field turned from verbal lists to visual arrays, because a visual display can be controlled with a precision that speech cannot. Steven Luck and Edward Vogel introduced the change-detection task that has dominated the field since: an array of coloured squares is shown briefly, removed, and after a blank interval a test array appears, either identical or differing in one item, and the observer judges same or different (Luck & Vogel, 1997). Because rehearsal cannot operate on so many colours in so short a time, accuracy as a function of set size yields a clean capacity estimate, and the answer converged on the same value found for attention and subitizing: about three to four items. Luck and Vogel further showed that a single object carrying several features, colour and orientation together, cost no more than an object with one feature, suggesting the stored unit is the integrated object rather than the feature.

Visual working memory also delivered a neural measure of capacity in the intact human brain. Vogel and Machizawa recorded a sustained, lateralised event-related potential during the retention interval, the contralateral delay activity, whose amplitude rose with the number of items held and then flattened at the individual's capacity limit, so that the waveform of a person holding two items differed from the same person holding four and predicted, across people, how many each could retain (Vogel & Machizawa, 2004). Here a scalp signal indexed the contents of the store trial by trial. The change-detection demonstration below reconstructs the task and estimates capacity from the responses using the standard formula, and the Worked Example computes that estimate by hand (Rouder et al., 2011).

Slots Versus Resources

The clean capacity estimates concealed a deep theoretical dispute that visual working memory then brought into the open: is capacity a matter of a few discrete slots, or of a single continuous resource? On the slot model, working memory holds a small fixed number of items, perhaps four, each stored at a fixed resolution; items beyond that number are simply not stored, and precision does not depend on how many are held. Weiwei Zhang and Steven Luck supplied strong evidence for this view by analysing not merely whether an item was remembered but how precisely: the distribution of errors in reporting a remembered colour showed a fixed resolution up to the capacity limit and pure guessing beyond it, exactly the signature of discrete, fixed-resolution slots (Zhang & Luck, 2008).

The resource model denies both the fixed number and the fixed resolution. On this account a limited pool of representational resource is distributed across all items in the display, so that the more items are held, the less resource each receives and the less precisely each is stored, with no hard item limit at all. Paul Bays and Masud Husain showed that the precision of visual memory declines continuously as set size grows, and does so even within the range where the slot model says resolution should be constant, and that resource can be allocated unequally, with more given to attended or recently cued items (Bays & Husain, 2008). A decade of increasingly sophisticated experiments has not decisively settled the question; the modern consensus, if there is one, is that the strict versions of both models are wrong and that capacity is best described as a variable, distributable resource with a strong item limit, a hybrid that the field is still refining (Ma, Husain, & Bays, 2014). The demonstration below contrasts the precision each model predicts as the number of remembered items rises.

Neural Basis

The physiological signature of working memory was discovered before the psychological construct was named. Recording from the prefrontal cortex of monkeys performing a delayed-response task, Joaquin Fuster and Garrett Alexander found neurons that fired steadily throughout the delay between a stimulus and the response it cued, bridging the gap during which the animal had to hold the information in mind (Fuster & Alexander, 1971). This persistent, stimulus-specific delay-period activity became the canonical neural correlate of maintenance. Shintaro Funahashi, Charles Bruce, and Patricia Goldman-Rakic sharpened the picture in an elegant experiment: prefrontal neurons in a spatial delayed-response task showed memory fields, firing during the delay only when the remembered target had appeared at a particular location, so that the population encoded which location was being held across the blank interval (Funahashi, Bruce, & Goldman-Rakic, 1989). Goldman-Rakic wove these findings into an influential circuit model in which recurrent excitation among prefrontal neurons sustains the representation and dopaminergic input tunes it, making the prefrontal cortex the seat of the mental workspace (Goldman-Rakic, 1995).

That tidy identification of maintenance with persistent prefrontal firing has since been complicated on two fronts. First, modern imaging and decoding work shows that the contents of working memory are not stored in the prefrontal cortex alone but are distributed across the sensory and parietal areas that process the material, with the prefrontal cortex supplying control rather than storage (D'Esposito & Postle, 2015). Second, the assumption that maintenance requires continuous firing has been challenged directly. Mark Stokes and others have argued that information can be held in activity-silent states, in short-term changes to synaptic weights that carry the memory through periods when the neurons are quiet and that can be read out by a probing input, a dynamic-coding framework in which the neural signature of a memory changes moment to moment rather than persisting unchanged (Stokes, 2015). Whether working memory is sustained activity, translated synaptic state, or both in alternation is among the liveliest questions in the field. The circuits that supply this control are concentrated in the prefrontal cortex.

Worked Example

Estimating visual working-memory capacity from a change-detection experiment is worth doing by hand, because the formula is simple, it is exactly what the change-detection demonstration computes, and its logic corrects a common error. Suppose an observer is tested with arrays of six items, so the set size N is 6. On change trials the observer correctly reports a change 90 per cent of the time, a hit rate H of 0.90; on no-change trials the observer wrongly reports a change 15 per cent of the time, a false-alarm rate F of 0.15. It is tempting to read the 90 per cent hit rate as evidence that nearly all six items were stored, but that ignores the guessing revealed by the false alarms. Cowan's formula corrects for it: capacity K equals the set size multiplied by the difference between the hit rate and the false-alarm rate, K = N(H − F) (Cowan, 2001).

Substituting, K = 6 × (0.90 − 0.15) = 6 × 0.75 = 4.5 items. The interpretation is exact: when the display holds six items, this observer retains information about four and a half of them, and the rest are guessed. The subtraction of the false-alarm rate is what turns a raw accuracy into a capacity, because an observer who pressed change on every trial would score a perfect hit rate while storing nothing, and the formula would correctly return K = 6 × (1 − 1) = 0. The single-probe and whole-display versions of the task use slightly different corrections, and choosing the right one for the design matters for comparing individuals, which is the methodological point that Rouder and colleagues make (Rouder et al., 2011). Moving the demonstration's set-size and performance controls recomputes K by this same equation.

Discussion

Working memory is where cognitive psychology keeps its clearest picture of the mind's limits and its least settled account of their cause. The clear picture is the convergence: a capacity of about four items surfaces in verbal recall, in visual change detection, in attentional tracking, and in the subitizing range, and a single scalp waveform tracks that capacity trial by trial, evidence that the limit is a real property of a real system and not an artefact of any one task (Cowan, 2001; Vogel & Machizawa, 2004). The unsettled account is what the system is. The multicomponent model describes it as a set of specialised buffers under executive control; the embedded-processes view describes it as the activated portion of long-term memory within the focus of attention; the resource models describe it as a divisible commodity with no fixed parts at all (Baddeley, 2003; Ma et al., 2014). These are not merely different vocabularies, because they make different predictions about precision, about whether items compete, and about what the neural data should show. What unites the modern field is the recognition that working memory is less a place where information sits than a state that information is held in, an attentional achievement that is continually renewed and continually under threat from the next incoming demand (Engle, 2002). Specifying the mechanism of that renewal, and reconciling the persistent-activity and activity-silent accounts of it, is the work that remains.

Commonly Confused With

Short-Term Memory

Short-term memory stores. Working memory stores and manipulates. Digit span forwards is short-term memory; digit span backwards is working memory, because the digits must be held and transformed at once. If the task has no transformation, it is not measuring working memory. The confusion is historical and lexical: the two were once one construct, and MeSH still indexes working memory as an entry term under Memory, Short-Term, which is why that mapping is a broadMatch and not an identity.

Common Misconceptions

Working memory holds seven items.

Miller's (1956) magical number seven described chunks in immediate memory, and Miller himself treated the recurrence of seven half-sceptically (Miller, 1956). When chunking is controlled, the pure capacity is closer to four items (Cowan, 2001). The seven survives because it was memorable, not because it was measured cleanly.

Working memory is just another name for short-term memory.

The two dissociate: simple digit span, a storage measure, predicts reading comprehension weakly, whereas complex span, which interleaves storage with processing, predicts it strongly (Daneman & Carpenter, 1980). Manipulation, not mere retention, is the defining feature, and it is what makes working-memory capacity predict reasoning (Engle, 2002).

Each remembered item is stored at the same fixed resolution.

The precision of a visual memory declines as more items are held, even within the range where a strict slot model predicts constant resolution, and resource can be shifted toward attended items (Bays & Husain, 2008). Capacity is not simply a count of fixed-quality slots; how well each item is stored depends on how many others share the load (Ma et al., 2014).

Glossary

Articulatory rehearsal.

The subvocal refreshing of speech-based material within the phonological loop, which offsets the decay of the phonological store and underlies the word-length effect.

Central executive.

The attentional control component of the multicomponent model, which allocates resources, coordinates the subsystems, and selects and inhibits, holding no information of its own.

Change detection.

A task in which an observer compares a memory array with a test array and judges whether one item has changed; the standard method for estimating visual working-memory capacity.

Chunk.

A unit of working-memory capacity formed by grouping several elements into one meaningful whole, so that recoding expands how much can be held without raising the item count.

Complex span.

A working-memory task that interleaves storage with an unrelated processing demand, such as reading span; a stronger predictor of reasoning and comprehension than simple span.

Contralateral delay activity.

A sustained event-related potential during the retention interval whose amplitude rises with the number of items held and plateaus at the individual's capacity limit.

Delay-period activity.

Persistent, stimulus-specific neural firing that spans the interval between a stimulus and a delayed response; the classical neural correlate of working-memory maintenance.

Episodic buffer.

The fourth component of the multicomponent model, a limited store that binds information from the subsystems and long-term memory into integrated, multidimensional episodes.

Memory field.

The property of a prefrontal neuron that fires during the delay only when the remembered item occupied a particular location, so the population encodes what is held.

Phonological loop.

The speech-based subsystem of working memory, comprising a phonological store that holds sound-coded material and an articulatory process that rehearses it.

Phonological similarity effect.

The poorer ordered recall of similar-sounding than dissimilar-sounding items, evidence that the phonological store codes material by sound rather than by appearance.

Resource model.

The account on which working memory is a continuous pool distributed across all items, so that precision falls as set size rises, with no fixed number of stored items.

Slot model.

The account on which working memory holds a small fixed number of items, each at a fixed resolution, with items beyond that number not stored at all.

Visuospatial sketchpad.

The subsystem of working memory that holds visual and spatial information for inspection and manipulation, dissociable from the verbal phonological loop.

Working memory.

The limited-capacity system that holds and manipulates information over intervals of seconds, distinguished from passive short-term storage by the operations it performs on its contents.

Key Researchers

Alan Baddeley (b. 1934). Professor of Psychology at the University of York; with Graham Hitch he proposed the multicomponent model of working memory and later added the episodic buffer. Faculty Page - ORCID - Google Scholar - Wikipedia

Nelson Cowan. Curators' Distinguished Professor of Psychological Sciences at the University of Missouri; he proposed the embedded-processes model and the estimate that pure capacity is about four items. Faculty Page - ORCID - Google Scholar - Wikipedia

Randall W. Engle. Professor of Psychology at the Georgia Institute of Technology; he developed the theory of working-memory capacity as executive attention and its link to fluid intelligence. Faculty Page - ORCID - Google Scholar - Wikipedia

Patricia S. Goldman-Rakic (1937-2003). Professor of Neuroscience at the Yale University School of Medicine; she mapped the prefrontal circuitry of working memory and characterised its persistent delay-period activity. Wikipedia - Wikidata

Steven J. Luck. Distinguished Professor of Psychology at the University of California, Davis; with Edward Vogel he established the capacity of visual working memory and its electrophysiological index. Faculty Page - ORCID - Google Scholar - Wikipedia

Edward K. Vogel. Professor of Psychology at the University of Chicago; with Steven Luck he measured visual working-memory capacity and identified the contralateral delay activity that indexes it. Faculty Page - ORCID - Google Scholar

Frequently Asked Questions

What is working memory?
Working memory is the limited-capacity system that holds and manipulates information over intervals of seconds, serving as the mental workspace in which the intermediate products of reasoning, comprehension, and problem solving are kept available while they are operated on (Baddeley, 2003).

What is the difference between working memory and short-term memory?
Short-term memory passively stores information, whereas working memory stores and manipulates it; repeating digits in order is a short-term task, while repeating them in reverse is a working-memory task because the items must be held and transformed at once (Daneman & Carpenter, 1980).

How many items can working memory hold?
When chunking is controlled, pure capacity is about four items rather than the more famous seven, a limit that recurs across verbal recall, visual storage, and attentional tracking (Cowan, 2001).

Why is the magical number seven considered wrong?
Miller's seven counted chunks and was measured without preventing spontaneous recoding, so it reflected chunking ability as much as raw capacity; controlled procedures that block chunking yield a limit closer to four (Cowan, 2001).

What are the components of Baddeley's model?
The model comprises a central executive that controls attention, a phonological loop for speech-based material, a visuospatial sketchpad for visual and spatial material, and an episodic buffer that binds information across them and with long-term memory (Baddeley, 2000).

How is visual working-memory capacity measured?
It is estimated from a change-detection task using Cowan's formula, in which capacity equals the set size times the difference between the hit rate and the false-alarm rate, correcting raw accuracy for guessing (Rouder et al., 2011).

Are memory slots real or is capacity a flexible resource?
The question is unresolved: error distributions support a small number of fixed-resolution slots, yet precision declines continuously with set size in a way that favours a divisible resource, and current models combine the two (Ma, Husain, & Bays, 2014).

Where in the brain is working memory held?
Maintenance was long identified with persistent firing in the prefrontal cortex, but current evidence distributes the stored contents across sensory and parietal areas and shows that information can also be held in transient, activity-silent synaptic states (D'Esposito & Postle, 2015).

References