Key facts
- What it is
- A limited system that holds information in an immediately usable state while operations are performed on it
- The defining feature
- Manipulation, not storage: short-term memory holds, working memory holds and works
- The standard model
- Baddeley and Hitch (1974): phonological loop, visuospatial sketchpad, central executive, plus the episodic buffer added in 2000
- Capacity
- About seven plus or minus two chunks with rehearsal allowed (Miller, 1956); closer to four items when rehearsal and chunking are prevented (Cowan, 2001)
- Duration
- Seconds, unless actively refreshed
- Neural basis
- The frontoparietal network: dorsolateral prefrontal cortex and posterior parietal cortex, with persistent delay-period firing
- Link to intelligence
- Substantial correlation with fluid intelligence, but the two are not the same thing and causation is unestablished
The distinction that makes the concept
Before 1974, psychology mostly spoke of short-term memory: a small store that held a few items for a few seconds before they faded or were displaced. The metaphor was a shelf. Things sat on it, briefly, and then fell off.
The problem with the shelf is that it does not do anything. A store that only stores cannot explain what happens when you multiply 17 by 4 in your head, or when you listen to a long sentence and have to keep its opening clause intact until the verb finally arrives, or when someone gives you directions with four turns in them and you have to hold turn three while executing turn one. In every one of those cases information is being held and transformed at the same time, and the transformation is not a separate step that happens afterwards. It happens on the held material, while it is held.
That is the whole insight, and it is worth stating as flatly as possible.
Short-term memory: the passive retention of information over a few seconds. Repeating a phone number straight back is a pure short-term memory task: what goes in comes out unchanged.
Working memory: the retention of information while operations are performed on it. Repeating the same phone number back in reverse is a working memory task, because the digits must be kept alive at the same time as they are being reordered. The load is the same; the demand is not.
This is why the two are measured with different tasks. Digit span forwards is a storage measure. Digit span backwards, or operation span, in which a person solves a small arithmetic problem after each word they are asked to remember, are working memory measures, because they force storage and processing to compete for the same limited resource. And they behave differently: backwards span is reliably shorter than forwards span, and it is the backwards and complex spans, not the simple ones, that predict reasoning ability. The manipulation is not an incidental extra. It is the part that matters.
Working memory, then, is not a bigger short-term memory. It is a different kind of thing: a workspace rather than a shelf. Everything else on this page follows from that.
Baddeley and Hitch: a workspace with parts
In 1974, Alan Baddeley and Graham Hitch published a chapter that reframed the field. Their argument was empirical and simple. If short-term memory were a single undifferentiated store, then loading it up should interfere with everything that uses it. So they loaded it up: they asked people to hold a string of digits in mind while simultaneously doing a reasoning task or a comprehension task. If there is one store, performance on the second task should collapse.
It did not collapse. It degraded, gracefully, and far less than a single-store account predicted. People could hold three digits and reason almost normally. That result is only explicable if the digits and the reasoning are drawing on partly separate resources. So Baddeley and Hitch proposed a system with parts.
The phonological loop
The system that holds speech-based information. It has two pieces: a passive phonological store, sometimes called the inner ear, which holds sound-based traces that decay in roughly two seconds; and an articulatory rehearsal process, the inner voice, which refreshes those traces by silently saying them again. This is why you keep a phone number alive by muttering it to yourself: you are literally re-entering it into the store before it fades.
The visuospatial sketchpad
The equivalent system for what things look like and where they are. It is what you use to picture the layout of your kitchen, to count the windows on the front of your house from memory, or to rotate a shape mentally. It has partly separable visual (appearance) and spatial (location) aspects, and it is disrupted by irrelevant visual input in the way the loop is disrupted by irrelevant speech.
The central executive
Not a store at all. The central executive is the attentional controller: it decides which subsystem gets resources, keeps the goal in view, suppresses irrelevant material, and coordinates the two slave systems when a task needs both. It is the boss, and, being a boss, it is the part of the model with the least concrete specification, a point Baddeley himself repeatedly conceded.
Two pieces of evidence tie the phonological loop down and are worth knowing because they show how a psychological model gets tested. The first is the word-length effect: people can hold more short words than long words, because rehearsal takes real time and long words take longer to say, so fewer of them can be refreshed before the store decays. The second is articulatory suppression: if you make people repeat "the, the, the" out loud while trying to remember a list, the rehearsal process is occupied and memory collapses. Both effects are exactly what the loop predicts and are hard to explain without something like it.
Why a model with parts earns its keep: it makes falsifiable predictions about interference. If the loop and the sketchpad are separate, then a verbal load should damage a verbal task much more than a spatial one, and vice versa. This dual-task logic is the workhorse of the whole field, and it is why the multicomponent model survived: it kept predicting which combinations of tasks would collide and which would not.
The buffer that had to be invented
The 1974 model had a hole in it, and Baddeley spent twenty-five years being honest about it before proposing a fix in 2000.
The hole is this. Suppose you are told to remember a red circle. The word "red" and the word "circle" could live in the phonological loop. The image could live in the sketchpad. But nothing in the original model says how the redness and the roundness get bound into one thing rather than two loose facts. Worse, suppose you are asked to remember a sentence rather than a list of unrelated words. People can hold roughly fifteen or sixteen words of a meaningful sentence, far beyond the capacity of any phonological store, because the sentence's meaning does the work. But meaning lives in long-term memory, and the 1974 model contained no route by which long-term knowledge could enter the workspace and combine with what was already there.
So the model as it stood could store, but it could not bind. That is a serious gap, because binding is what experience is made of: you do not perceive a colour and a shape and a location as three unrelated registrations, you perceive an object.
Episodic buffer (Baddeley, 2000): a limited-capacity, multidimensional store that binds information from the phonological loop, the visuospatial sketchpad and long-term memory into integrated episodes. It is the component that lets working memory hold a scene rather than a list, and it is the interface through which existing knowledge reaches the current workspace.
The buffer is called episodic because what it holds are episodes: bound, coherent chunks of experience with several attributes at once. It is assumed to be controlled by the central executive and to be accessible to conscious awareness, which makes it, in the model's own terms, the place where the contents of working memory are actually experienced.
It is fair to note that the episodic buffer is the least tidy part of the model, and that Baddeley proposed it as a solution to a problem rather than as something demanded by a specific experiment. That is not a criticism. It is how theories usually get repaired: a phenomenon appears that the existing components cannot produce, and the model has to grow a component that can.
Seven, or four: the most misquoted number in psychology
In 1956, George Miller published a paper in Psychological Review called "The magical number seven, plus or minus two". Nearly everyone has heard of it. Almost everyone has it wrong.
What people think it says: you can hold about seven things in mind at once.
What it actually says: the span of immediate memory is limited to about seven chunks, and a chunk is not an item in the world, it is a unit of meaning to the person doing the remembering. Miller's central point was precisely that the limit is not on information in any absolute sense, because the amount of raw information inside a chunk can vary enormously. He was drawing a distinction, not stating a hard constant, and his title was written with a distinctly light touch: he opens by complaining, half in jest, of being persecuted by an integer.
Chunk: a collection of elements bound into a single unit by knowledge already held in long-term memory. "IBM" is one chunk to an English speaker who knows the company and three unrelated letters to someone who does not. The number 1066 is one chunk to a British schoolchild and four digits to everyone else. Capacity is counted in chunks, which is why capacity measured in items appears to change with what you happen to know.
This immediately explains why the seven-item claim will not survive contact with data. If a chunk can be a letter or a whole familiar acronym, then "how many items can you hold" has no fixed answer, because the size of the container depends on the contents of your long-term memory. Any experiment that lets people rehearse and chunk is measuring their knowledge as much as their capacity.
This is exactly the problem Nelson Cowan set out to solve in a 2001 review in Behavioral and Brain Sciences. His strategy was to gather the studies that had blocked the escape routes: designs that prevented rehearsal (by occupying the articulatory process), prevented chunking (by using unrelated material, or by presenting information too fast to group), and required all the information to be held at once. Under those conditions, the estimates converged, and they converged not on seven but on about four. Cowan's argument is that four is the real capacity of what he calls the focus of attention, the set of items that can be held simultaneously in an active, immediately available state.
The two numbers are not in conflict, and treating them as rival claims is the standard error. They are answers to different questions. Miller measured performance in a task where people were free to use every trick available: rehearse, group, recode, lean on what they already knew. Seven is what a knowledgeable, rehearsing human can juggle. Cowan measured what is left when those tricks are taken away. Four is the capacity of the raw workspace. The gap between four and seven is not noise; it is the size of the contribution that rehearsal and long-term knowledge make to what looks like memory capacity.
Chunking, and why chess masters are not really remembering more
If chunks are the unit, then improving your memory is largely a matter of building better chunks, and the classic demonstration comes from chess.
Show a chess master a position from a real game for a few seconds and take it away, and they can reconstruct it with startling accuracy, far beyond what a novice manages. This looks like superhuman memory, and for a long time it was taken as evidence of one. It is not. The critical control is to show them a board with the pieces scattered at random, in configurations that could not arise in a real game. On random boards, the master's advantage largely evaporates: they perform close to the level of a novice.
The explanation is chunking, and it is decisive. The master is not holding twenty-five pieces. They are holding a handful of chunks: a castled king position, a familiar pawn structure, a known attacking formation. Years of study have compiled those recurring configurations into single units in long-term memory, so that what a novice must encode as twenty-five separate facts the master encodes as five or six meaningful ones. Randomise the board and the configurations no longer correspond to anything in that long-term store, so there is nothing to chunk with, and capacity falls back to the ordinary human handful.
The general lesson, which applies well beyond chess: expertise does not enlarge working memory. It supplies better chunks, so that the same small workspace can hold more of the world. This is why domain knowledge makes people look as though they think faster, and why an expert dropped outside their domain is as ordinary as anyone else. It is also why "memory tricks" work: mnemonics are chunking devices, imposing structure on material that had none.
The same principle explains why you can hold a sentence far longer than a random word list of the same length, why a familiar telephone number is easier than an unfamiliar one, and why the string 1-4-9-2-1-7-7-6 is trivial for someone who sees two dates in it and awkward for someone who sees eight digits. Long-term memory is not a separate department from working memory. It is the thing that determines how much working memory you effectively have. For how those durable stores are built in the first place, see memory consolidation and types of memory.
Watching a memory being held
Working memory is one of the few cognitive constructs for which there is a plausible candidate for the physical act itself, observable in single neurons.
The paradigm is the delayed-response task. An animal sees a cue, the cue disappears, a delay of several seconds follows in which nothing at all is present, and then the animal must respond on the basis of the cue that is no longer there. During that empty delay, the information exists nowhere in the outside world. If the animal is going to answer correctly, the brain must be maintaining it.
And in the dorsolateral prefrontal cortex, some neurons do exactly what you would want them to do: they fire steadily throughout the delay. They begin firing when the cue appears, keep firing while the screen is blank, and stop when the response is made. Many of them are selective, firing during the delay only for a cue in a particular location or of a particular kind. This persistent delay-period activity is as close as neuroscience has come to watching a memory being held rather than merely stored, and it has been the anchor of the field for decades.
Persistent (delay-period) activity: sustained firing of a neuron during an interval in which the item it codes for is absent from the environment. Because the stimulus is gone, the activity cannot be a response to it; it is generally interpreted as the maintenance of an internal representation.
The anatomy is a network, not a spot. Human imaging consistently implicates the frontoparietal network: the dorsolateral prefrontal cortex together with the posterior parietal cortex, the same network that supports executive function and attention. The rough division of labour that most accounts endorse is that parietal cortex carries much of the content, the stored representation itself, while prefrontal cortex does the controlling, protecting the representation against distraction and directing what enters and leaves. Sensory cortex is involved too: activity in visual areas during a delay carries information about what is being held, which suggests that maintaining a visual item partly means keeping its sensory representation alive. Dopaminergic neuromodulation appears to help set the gate, biasing whether a representation is stably held or flexibly updated, which is why prefrontal dopamine levels affect working memory performance.
Honesty requires a caveat, and it is a substantial one. Persistent activity is not the whole story, and the assumption that it is the storage mechanism has come under serious challenge. Modern recordings show that delay activity is often far less steady than the classic picture implies: it fluctuates, and in some analyses the population code appears to go quiet and then return. This has prompted activity-silent accounts, in which information is held for short periods in rapid, transient changes to synaptic strength rather than in ongoing firing, and is read back out when needed. On such an account, a memory can persist through a period in which no neuron is visibly carrying it. The debate is live and unresolved. What is not in dispute is that the frontoparietal network is where the maintenance happens; what is in dispute is precisely how the holding is physically implemented.
Why this page sits on an intelligence site
Working memory is not merely one cognitive function among many. It is the single cognitive capacity most closely tied to measured reasoning ability, and that is why it belongs here rather than as a footnote.
The empirical claim, stated carefully: measures of working memory capacity, particularly complex span tasks that force storage and processing to compete, correlate substantially with measures of fluid intelligence, the ability to reason about novel problems. The size of the correlation depends heavily on how both constructs are measured; in latent-variable studies, which combine several tasks per construct to strip out the quirks of any one of them, the reported relationship is commonly around 0.5 or higher, and some analyses have found it higher still. Few relationships in differential psychology are this robust across laboratories and methods.
The reason it should be true is not mysterious, and you can feel it working. Consider what a matrix reasoning item, the archetypal fluid intelligence test, actually demands. You must hold a candidate rule in mind. You must apply it across a row while keeping it intact. You must remember which candidates you have already ruled out, so that you do not go round in circles. You must resist the pull of an answer that looks superficially right. Every one of those steps is a demand on a small workspace. Reasoning is, mechanically, the manipulation of held premises. Give someone a bigger and better-protected workspace and you would expect them to reason better, for the same reason that a bigger desk helps if the job involves comparing documents.
That is the plausible story. Now the disciplined part, which matters more.
What the correlation does not license: first, correlation is not identity. Even a correlation of 0.5 leaves most of the variance in fluid intelligence unaccounted for by working memory, and researchers who have argued the two constructs are nearly the same thing at the latent level are making a strong and contested claim, not reporting a settled finding. Second, the causal direction is not established. Better maintenance might produce better reasoning; better reasoning might produce more efficient use of a workspace by generating better chunks and better strategies; both might depend on a third factor, such as the integrity or efficiency of the frontoparietal network. Correlational data cannot distinguish these, and the training evidence, which is the natural way to test the causal claim, has been discouraging. Third, and following directly: raising working memory scores does not reliably raise general ability, which is what you would predict if the relationship were causal and simple. It is not.
The right summary is therefore precise rather than exciting. Working memory capacity is one of the best single-construct correlates of fluid reasoning we have, and it gives a mechanistic account of why some minds handle novel problems better. It is not intelligence, it does not replace an intelligence test, and it is not a lever you can pull to raise one.
What people get wrong
Myth: you can hold seven items in working memory, plus or minus two.
Fact: Miller's 1956 figure was about chunks, not items, and a chunk is defined by the knowledge of the person doing the remembering. Because chunking and rehearsal inflate apparent capacity, the number tells you as much about what someone knows as about the size of their workspace. When rehearsal and chunking are experimentally prevented, capacity estimates converge on about four items (Cowan, 2001). Quoting seven as a hard limit on items is a misreading of the paper it comes from, and Miller himself framed the number with a good deal more irony than his citers have.
Myth: you can train working memory and raise your IQ.
Fact: training on n-back and similar tasks reliably improves performance on the trained task, and often on tasks that closely resemble it. That is near transfer, and it proves very little, because getting better at a task you have practised is what practice does. Far transfer, an improvement in general reasoning ability, is weak and unreliable. The largest early trial, by Owen and colleagues in Nature in 2010, trained over eleven thousand participants online for six weeks and found gains on the trained games and no meaningful benefit on untrained measures of general cognitive ability. The correlation between working memory and fluid intelligence is real, but a correlation is not a lever.
Myth: working memory is just short-term memory with a bigger capacity.
Fact: it is defined by manipulation, not by size. A system that held twenty items but could do nothing with them would still not be working memory. This is why forwards digit span, a pure storage measure, is a poor predictor of reasoning, while backwards span and complex span, which force processing and storage to compete, predict it well. The clue is in the name: the memory has to work.
Myth: a good memory means a large working memory.
Fact: they are largely different systems. People with extraordinary long-term recall, including memory-competition champions, generally have ordinary working memory capacity; what they have is a trained set of chunking and encoding strategies. Conversely, dense amnesia can leave immediate span intact: patients who cannot form new long-term memories at all may still repeat a string of digits back normally, because the workspace is not the archive. See types of memory for how these systems separate.
Where you feel the limit
Working memory is unusual among cognitive constructs in that its limit is not an abstraction. You bump into it several times a day, and you know exactly what it feels like when you do.
Mental arithmetic. Multiplying 17 by 24 without paper is the purest working memory task in ordinary life. You have to hold the partial products while computing the next one, and the reason it feels like carrying something heavy is that it is: each partial product occupies part of a small, decaying store. Most people fail not because they cannot do the multiplication but because a partial product falls out of the workspace before it can be used.
Following a spoken direction. "Take the second left, go past the church, then it is the third turning on the right after the roundabout." Reading that is easy. Being told it, once, while driving, and executing it, is not. The instructions must be held, in order, while attention is also being spent on the road, and by the roundabout the earlier items have decayed.
Understanding a long sentence. Comprehension of a sentence whose subject is a long way from its verb requires holding the opening of the sentence intact until the structure resolves. This is why sentences with deeply embedded clauses feel effortful and why they collapse into incomprehensibility past a certain depth. Language processing runs on working memory, which is why language and the brain and this page keep meeting.
Walking into a room and forgetting why. The goal was being held in an active state, and something displaced it: a thought, a doorway, a person speaking. The information was never encoded into long-term memory, because it did not need to be; it only had to survive for fifteen seconds. When it did not, there is nothing to retrieve, which is why the feeling is one of blankness rather than of a memory you cannot reach.
Losing the thread when interrupted. The reason an interruption is expensive out of all proportion to its length is that it does not merely pause the task, it evicts the task's contents from the workspace. Reconstructing them takes far longer than the interruption did.
Two practical consequences follow, and they are the honest ones. First, since capacity is fixed but chunk size is not, the way to work around the limit is to change the material, not the mind: write things down, group them, give them structure, and let long-term knowledge do the compression. Second, working memory is unusually sensitive to state. It degrades with poor sleep, with acute stress, and with anxiety, all of which occupy the same limited attentional resource. Anyone who has tried to do arithmetic while worrying has run the experiment.
Sources
- Baddeley AD, Hitch GJ. Working memory. In: Bower GA, ed. The Psychology of Learning and Motivation, vol 8. Academic Press; 1974:47-89.
- Baddeley A. The episodic buffer: a new component of working memory? Trends in Cognitive Sciences. 2000;4(11):417-423.
- Cowan N. The magical number 4 in short-term memory: a reconsideration of mental storage capacity. Behavioral and Brain Sciences. 2001;24(1):87-114.
- Miller GA. The magical number seven, plus or minus two: some limits on our capacity for processing information. Psychological Review. 1956;63(2):81-97.
- Owen AM, Hampshire A, Grahn JA, et al. Putting brain training to the test. Nature. 2010;465(7299):775-778.
- Kandel ER, Koester JD, Mack SH, Siegelbaum SA. Principles of Neural Science. 6th ed. McGraw-Hill; 2021.
This page is an educational reference. It is not medical advice and does not diagnose or treat any condition.