Working Memory

Abstract

Working memory is the limited-capacity system that holds and manipulates information over intervals of seconds, supporting cognition in progress rather than storing knowledge for later retrieval. This article traces the construct from the unitary short-term store through the multicomponent model of Baddeley and Hitch to the embedded-processes account. Its central controversy is the nature of the capacity limit: whether storage is bounded by a fixed number of discrete slots or by a continuous resource spread across items. Three interactive demonstrations let the reader examine the word-length effect, estimate capacity from change detection, and pit the two accounts against each other.

Keywords: working memory, capacity limit, change detection

Working memory is the system that holds a limited amount of information in an accessible state and manipulates it in the service of ongoing thought (Baddeley & Hitch, 1974; Baddeley, 1992). It is one of the principal forms of human memory, the fast-access mental workspace that operates alongside the long-term store rather than within it. It differs from the older notion of short-term memory in emphasizing not just brief storage but active use: the arithmetic, reasoning, comprehension, and planning that require several pieces of information to be kept in mind and combined. Working memory matters to cognitive psychology out of proportion to its tiny size, for three reasons. Its architecture, a set of specialized buffers coordinated by a control process, is one of the most influential theories of the structure of the mind. Its capacity limit, famously a mere handful of items, is among the most reliable individual differences in psychology and predicts reasoning ability, reading comprehension, and fluid intelligence. And the question of what that limit actually is, a fixed number of discrete slots or a continuous resource spread ever thinner, remains one of the liveliest open debates in the field. The sections below build the model component by component, examine the capacity limit and how it is measured, connect working memory to attention and intelligence, and lay out the slots-versus-resources controversy and its roots in the brain.

From Short-Term Store to Working Memory

The modern concept grew out of, and then broke with, the idea of a unitary short-term store. George Miller's 1956 essay observed that the span of immediate memory hovers around seven items across wildly different materials, a regularity so robust he called it a magical number, while noting that the true unit is the chunk, a meaningful group whose formation lets people pack far more raw detail into the same few slots (Miller, 1956). For two decades short-term memory was pictured as a single passive buffer that held such items briefly before they were either rehearsed into long-term memory or lost. Alan Baddeley and Graham Hitch overturned that picture in 1974 by showing that people can hold a string of digits and simultaneously reason or comprehend, with only modest interference, which a single shared store cannot easily explain (Baddeley & Hitch, 1974). They proposed instead a working memory of several interacting parts, and renamed the system to stress that its job is to work on information, not merely to hold it. The shift from store to workspace reframed the whole topic: the interesting questions became how the parts divide labor and how a central process allocates their limited resources.

The Multicomponent Model

Baddeley and Hitch's model, refined over the following decades, is built from a controlling attentional process and a set of subsidiary storage buffers (Baddeley, 1992; Baddeley, 2003). The central executive is the coordinator: a limited attentional controller that directs focus, switches between tasks, retrieves from long-term memory, and allocates the storage systems, but that stores little itself. Under it sit two domain-specific buffers. The phonological loop holds speech-based material through a brief phonological store that fades in about two seconds and an articulatory rehearsal process that refreshes it by inner speech, an arrangement whose signature is the word-length effect: memory span is shorter for words that take longer to say, because fewer of them fit in the rehearsal window. The loop is not a laboratory curiosity; Baddeley, Gathercole, and Papagno argued it evolved as a language-learning device, the store that holds an unfamiliar sound pattern long enough for a durable representation of a new word to form (Baddeley, Gathercole, & Papagno, 1998). The visuospatial sketchpad is the parallel buffer for visual and spatial material, holding the appearance and arrangement of objects and supporting mental imagery. In 2000 Baddeley added a fourth component, the episodic buffer, a limited store that binds information from the loop, the sketchpad, and long-term memory into unified, multidimensional episodes, filling a gap the original three parts could not: how the separate streams are integrated into a single conscious scene (Baddeley, 2000). Figure 1 shows the four components and their relations. The demonstration below reveals how the word-length effect exposes the timing of the phonological loop.

The Capacity Limit

The single most famous fact about working memory is that it holds very little, but the exact number has been repeatedly revised downward as measurement improved. Miller's seven plus or minus two conflated pure storage with the gains from chunking and rehearsal (Miller, 1956). When those aids are stripped away, by preventing rehearsal or by using materials that cannot be grouped, the underlying limit is closer to four. Nelson Cowan's influential reconsideration assembled converging evidence, from memory tasks, attention tasks, and enumeration, that the central capacity of young adults is about three to five chunks, a limit he later summarized as a magical number four (Cowan, 2001; Cowan, 2010). This same limit of about four appears when observers try to hold objects in mind at a glance, which is why the enumeration limit of subitizing and its shared capacity of about four is treated as the same constraint seen from the side of vision. Measuring the limit reliably is a craft in itself: the operation span, reading span, and change-detection tasks that dominate the field each try to isolate storage from processing, and their methodology has been codified precisely because small procedural choices move the estimate (Conway, Kane, Bunting, Hambrick, Wilhelm, & Engle, 2005). Table 1 collects the main capacity estimates and the assumptions behind them. The demonstration below computes a capacity estimate from hits and false alarms the way researchers do.

Table 1
Estimates of Working Memory Capacity by Method and Assumption

Note. Estimates fall as measurement more strictly separates storage from rehearsal and chunking. Sources are the immediate-span tradition (Miller, 1956), the reconsidered pure limit (Cowan, 2001), the visual change-detection estimate (Luck & Vogel, 1997), and the resource account that denies a fixed number (Ma, Husain, & Bays, 2014).

Working Memory and Attention

What makes working memory capacity so consequential is that it is less about storage than about the control of attention. Randall Engle and colleagues argued that the central executive is essentially executive attention, the ability to keep goal-relevant information active and to resist distraction, and that individual differences in working memory capacity are differences in this attentional control rather than in the size of a passive store (Engle, 2002). The evidence is that people with high and low working memory span differ not only on memory tasks but on tasks with minimal memory load that demand attention, such as resisting a reflexive glance or naming the ink color of a conflicting word. Most strikingly, working memory capacity is one of the strongest known predictors of fluid intelligence, the ability to reason and solve novel problems. Using latent-variable methods to separate the shared core of many tasks from their surface details, Engle, Tuholski, Laughlin, and Conway showed that working memory and short-term memory are separable, and that it is the working memory component, the part that involves controlled attention rather than mere storage, that carries the strong relationship to fluid intelligence (Engle, Tuholski, Laughlin, & Conway, 1999). Working memory, on this view, is a bottleneck on thought itself: how much a person can hold and protect in mind sets a ceiling on how complex a problem they can reason through.

Visual Working Memory

The visual side of the system has become the main testing ground for theories of capacity, because visual arrays can be controlled precisely and probed with change detection. Steven Luck and Edward Vogel had observers hold arrays of colored, oriented bars and then detect a change, and found a sharp limit of about four items, whether each item was defined by a single feature or by a conjunction of several, suggesting that the unit of storage is the integrated object rather than the individual feature (Luck & Vogel, 1997). The limit is not merely behavioral. Vogel and Machizawa discovered an electrophysiological signature, a sustained contralateral brain response during the retention interval whose amplitude rises with the number of items held and then flattens once an individual's capacity is reached, providing a direct neural readout of how much is currently in visual working memory and tracking each person's behavioral limit (Vogel & Machizawa, 2004). That a person's private capacity can be read from a brain wave made visual working memory a favored model system, and it sharpened the central theoretical question: is that limit a matter of counting discrete objects, or of dividing a fixed pool of quality among however many objects there are?

Slots Versus Resources

The deepest current debate is over the nature of the limit itself. The slot model holds that working memory consists of a small number of discrete, fixed-resolution slots, perhaps four, each storing one object at a set precision; items that exceed the slots are simply not stored, so performance reflects the probability that a probed item happened to occupy a slot. Weiwei Zhang and Steven Luck supported this view by showing that the precision of stored colors stays constant whether observers hold few or many items, while the probability of having stored the probed item falls, exactly as a fixed number of fixed-quality slots predicts (Zhang & Luck, 2008). The resource model holds the opposite: working memory is a continuous quantity shared among all items, so that adding items does not push some out but spreads the resource thinner, degrading the precision of every item smoothly with no hard limit on number. Paul Bays and Masud Husain reported that recall error grows continuously with set size and that resource can be shifted dynamically toward attended or behaviorally relevant items, evidence for a flexible pool rather than rigid slots (Bays & Husain, 2008). A later synthesis argued that the accumulated behavioral and neural evidence favors reconceiving working memory as a distributed, flexible resource in which quality, not a fixed quantity, is the fundamental currency (Ma, Husain, & Bays, 2014). The debate is not fully settled, and hybrid models allowing a variable number of unequal representations now occupy much of the middle ground. The demonstration below contrasts the two accounts directly, showing how each predicts precision to change as the number of items grows.

The Brain's Working Memory

The neural basis of working memory has itself shifted from a simple to a subtle picture. The classic view, from single-neuron recordings in the prefrontal cortex, was that information is held by persistent, elevated firing that bridges the delay between a stimulus and its use, with the prefrontal cortex acting as the seat of the store. The modern cognitive neuroscience of working memory complicates this in two ways (D'Esposito & Postle, 2015). First, maintenance appears to be distributed rather than localized: the content of working memory can be decoded from the same sensory and association areas that process the material in the first place, so holding a color in mind reengages visual cortex, and the prefrontal cortex looks more like a source of top-down control than a storehouse. Second, persistent firing is not the only mechanism; information can also be held in rapidly changing patterns and in short-term synaptic changes, so-called activity-silent states, from which it can be reactivated when needed. This dovetails with the psychological evidence, from the contralateral delay signal that indexes visual load (Vogel & Machizawa, 2004) to the resource view of capacity, that working memory is less a dedicated container than a controlled, temporary state imposed on the brain's ordinary machinery.

Worked Example

Because items that exceed capacity are simply guessed at, raw accuracy on a memory test overstates how much was truly held; the standard correction is Cowan's formula for the number of items in memory, K equals the set size N multiplied by the hit rate minus the false-alarm rate (Cowan, 2001). Consider the change-detection task in the demonstration above. An observer sees an array of N equals 8 colored squares, the array disappears, and a single square returns either unchanged or in a new color; the observer says whether it changed. Suppose that across many trials the observer correctly reports a change when one occurred on 75 percent of trials, a hit rate of 0.75, but also falsely reports a change when none occurred on 25 percent of trials, a false-alarm rate of 0.25. The corrected estimate is K equals 8 multiplied by the quantity 0.75 minus 0.25, which is 8 multiplied by 0.50, which equals 4.0 items. The false-alarm subtraction is what removes the inflation from guessing: an observer who simply guessed change on half of all trials would score 0.50 hits and 0.50 false alarms, giving K equals 8 multiplied by 0, that is 0 items, correctly recording that nothing was actually retained. Now shrink the array to N equals 4. An observer whose true capacity is four can store the whole array, approaching a hit rate of 1.0 and a false-alarm rate near 0, so K equals 4 multiplied by the quantity 1.0 minus 0, which is 4.0 items. The estimate lands near four from both directions, which is exactly why the change-detection task, corrected for guessing, became a workhorse for measuring a capacity that hovers stubbornly around four.

Discussion

Working memory matters first because it is a theory of the architecture of thought. The multicomponent model gave psychology a durable, testable account of how the mind holds and juggles information, one that has organized research for half a century and continues to be revised rather than replaced (Baddeley, 2003). It matters second because its capacity limit is one of the most consequential individual differences known: the amount a person can hold and control in mind predicts reasoning, reading comprehension, and fluid intelligence, making working memory a central variable in the science of why people differ in how well they think (Engle, Tuholski, Laughlin, & Conway, 1999). It matters third because the unresolved question of what the limit is, discrete slots or a shared resource, is not a technicality but a fork in our understanding of how the mind represents the world, with consequences for models of attention, perception, and decision-making (Ma, Husain, & Bays, 2014). And it matters in practice, because working memory limits shape everything from how instructions should be designed to how cognitive difficulties in aging, attention disorders, and schizophrenia are understood. The simple fact that the mind can hold only a few things at once turns out to be one of the deepest constraints on human cognition.

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 the same time. If a task carries no transformation, it is not measuring working memory. The two are historically the same construct, and older textbooks use the terms interchangeably; MeSH still indexes working memory under Memory, Short-Term (D008570), which is why that mapping is a broadMatch and not an exactMatch.

Common Misconceptions

Working memory holds seven items.

Miller's magical number seven described chunks in immediate serial recall, and Miller himself treated the figure lightly. When rehearsal and chunking are controlled, the pure limit is closer to four (Cowan, 2001). The seven survives because it was memorable, not because it isolated storage from strategy.

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

The two are separable. Latent-variable work shows passive short-term storage and working memory load onto distinct factors, and it is the working memory factor, the one carrying controlled attention rather than mere storage, that predicts fluid intelligence (Engle, Tuholski, Laughlin, & Conway, 1999). Short-term memory is the storage component without the executive control.

Working memory is stored in the prefrontal cortex.

Its contents are decoded from the same sensory and association areas that process the material, while the prefrontal cortex supplies top-down control rather than acting as the store; maintenance also rests partly on activity-silent synaptic traces, not on persistent firing alone (D'Esposito & Postle, 2015). The prefrontal store is a classic picture the evidence has outgrown.

Glossary

Capacity limit.

The small number of items or chunks, about four when strategy is controlled, that working memory can hold at once.

Central executive.

The attentional controller of working memory that directs focus, switches tasks, and allocates the storage buffers.

Change detection.

A task in which observers judge whether a probed item in a remembered array has changed, used to measure capacity.

Chunk.

A meaningful unit formed by grouping; the true unit of the capacity limit, allowing more raw detail per slot.

Contralateral delay activity.

A sustained brain response during retention whose amplitude tracks the number of items in visual working memory.

Cowan's K.

An estimate of items held, computed as set size times the hit rate minus the false-alarm rate, correcting for guessing.

Episodic buffer.

A limited store that binds material from the other buffers and long-term memory into unified, multidimensional episodes.

Executive attention.

The controlled maintenance of goal-relevant information and resistance to distraction, central to working memory capacity.

Fluid intelligence.

The ability to reason and solve novel problems, strongly predicted by working memory capacity.

Phonological loop.

The buffer that holds speech-based material through a fading store and articulatory rehearsal by inner speech.

Resource model.

The view that working memory is a continuous quantity shared among items, so precision falls as their number rises.

Short-term memory.

The brief passive storage of a small amount of information; the storage aspect of working memory without the control.

Slot model.

The view that working memory holds a fixed number of discrete items at fixed resolution, with excess items not stored.

Visuospatial sketchpad.

The buffer that holds visual and spatial material and supports mental imagery.

Word-length effect.

The finding that memory span is smaller for longer-to-pronounce words, revealing the timing of the phonological loop.

Working memory.

The limited-capacity system that holds and manipulates information in an accessible state for ongoing cognition.

Key Researchers

Alan Baddeley. Emeritus 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, providing the field's dominant framework for the structure of temporary memory.
Faculty Page · ORCID · Google Scholar · Wikipedia

Graham Hitch. Emeritus Professor of Psychology at the University of York; co-author with Baddeley of the original 1974 working memory model, and a continuing contributor to accounts of the phonological loop and serial order in short-term memory.
Faculty Page · Google Scholar · Wikipedia

Nelson Cowan. Curators' Distinguished Professor of Psychological Sciences at the University of Missouri; his embedded-processes model and his reconsideration of the capacity limit established the modern estimate of about four chunks.
Faculty Page · ORCID · Google Scholar · Wikipedia

Randall Engle. Professor of Psychology at the Georgia Institute of Technology; he reframed working memory capacity as executive attention and established its strong relationship to fluid intelligence through latent-variable studies of individual differences.
Faculty Page · ORCID · Google Scholar · Wikipedia

George A. Miller (1920–2012). Professor of Psychology at Princeton University and a founder of cognitive psychology; his 1956 paper on the magical number seven put the limits of immediate memory at the center of the field and introduced the chunk as its true unit.
Wikipedia

Steven Luck. Distinguished Professor of Psychology at the University of California, Davis; his change-detection studies established the object-based capacity limit of visual working memory and, with Weiwei Zhang, the discrete-slot account of that limit.
Faculty Page · ORCID · Google Scholar

Masud Husain. Professor of Neurology and Cognitive Neuroscience at the University of Oxford; with Paul Bays he developed the resource account of working memory, showing that memory precision degrades continuously and that resource can be allocated flexibly across items.
Faculty Page · ORCID · Wikipedia

Frequently Asked Questions

What is working memory?
Working memory is the system that holds a small amount of information in an accessible state and manipulates it in the service of ongoing thought, such as reasoning, comprehension, or mental arithmetic. It is distinguished from long-term memory by its tiny capacity and rapid loss, and from simple short-term storage by its emphasis on active control (Baddeley, 1992).

How is working memory different from short-term memory?
Short-term memory refers to the brief, passive storage of a small amount of information, whereas working memory adds the active manipulation and control of that information by an attentional system. In practice the two are separable but correlated, and it is the working memory component, involving controlled attention, that predicts reasoning ability (Engle, Tuholski, Laughlin, & Conway, 1999).

How many things can working memory hold?
When rehearsal and grouping are prevented, the limit is about four chunks in young adults, lower than the famous seven once strategy is stripped away. A chunk can be a single item or a meaningful group, so the effective amount held depends on how well the material can be organized (Cowan, 2001).

What are the components of working memory?
In Baddeley and Hitch's model, a central executive controls attention and coordinates three storage systems: a phonological loop for verbal material, a visuospatial sketchpad for visual and spatial material, and an episodic buffer that binds their contents with long-term memory into unified episodes (Baddeley, 2000).

What is the word-length effect?
The word-length effect is the finding that memory span is smaller for words that take longer to say. It arises because the phonological loop refreshes material by inner speech within a limited time window, so fewer long words than short words fit before the store fades, revealing the timing basis of verbal working memory (Baddeley, Gathercole, & Papagno, 1998).

Is working memory related to intelligence?
Yes. Working memory capacity is one of the strongest predictors of fluid intelligence, the ability to reason and solve novel problems. The link is thought to reflect shared reliance on controlled attention rather than on storage alone, so people who can better hold and protect information in mind tend to reason more effectively (Engle, 2002).

What is the slots-versus-resources debate?
It is the question of whether working memory holds a fixed number of discrete items at fixed precision, the slot view, or a continuous resource spread among all items so that precision falls as their number rises, the resource view. Evidence exists for both, and many researchers now favor hybrid models with a variable number of unequal representations (Ma, Husain, & Bays, 2014).

Where in the brain is working memory?
Rather than residing in a single store, working memory appears to be a temporary state distributed across the same sensory and association areas that process the material, with the prefrontal cortex providing top-down control. Information can be held both by sustained neural firing and by activity-silent synaptic traces that are reactivated when needed (D'Esposito & Postle, 2015).

References