Working Memory

011 min

Where working memory shows up

An illustrative case, using a role rather than a person. A designer is in a review, presenting three screens. A stakeholder interrupts with a question about a fourth screen that is not on the wall, then asks how the change would affect the pricing tier, then mentions a customer name. The designer answers the last question, pauses, and asks what the first one was. They have not lost concentration. They have run out of room.

Change one input and run it again. The same three questions arrive, but each is written on the shared document as it is asked. The designer answers all three, in order, and catches that the second and third contradict each other.

Nothing changed about the questions or about the designer. What changed was whether the questions had to be held in mind while being answered.

The moment the limit binds is specific and observable. It is the point where new information arrives while existing information still has to be kept available. Reading a long sentence, comparing two prices, following a set of directions, holding a bug's reproduction steps while reading the code that causes it — all of these are the same operation, and all of them fail in the same way at roughly the same point.

022 min

Why working memory is limited

Working memory is not one thing. Alan Baddeley and Graham Hitch proposed in 1974 that it is a system with separate parts, and that model still organises the field.

The components

One part handles verbal material, holding sound-based information for a short time and refreshing it by silent repetition. A second part handles visual and spatial material. A third part, the central executive, directs attention between them and decides what gets refreshed. Baddeley added a fourth component in 2000, the episodic buffer, to explain how information from the separate parts gets combined into single episodes and linked to long-term memory.

The reason a system needs parts at all is a finding anyone can reproduce. Holding a phone number while reading a sentence is hard. Holding a phone number while looking at a picture is much easier. If there were one pool of capacity, both would interfere equally.

The capacity number, and the conditions on it

Capacity is counted in chunks, which are units that have become meaningful through learning. The letters I, B and M are three chunks to someone who does not know the company and one chunk to someone who does. This is why capacity in chunks stays roughly fixed while the amount of raw material it holds varies enormously with expertise.

Nelson Cowan's review of the evidence puts the limit of the focus of attention at three to four chunks. The conditions attached to that number matter as much as the number:

• Rehearsal has to be prevented, by rapid presentation or by an unpredictable end point
• Items must not be groupable into larger meaningful units
• The task must require attention-based storage rather than leaning on passive sensory traces

When those conditions hold, Cowan reports the limit is about three and a half chunks across learning conditions in word-pairing work, and between three and four items for visual or mixed stimuli.

When those conditions do not hold — which is most of ordinary life — people appear to hold far more, because they are quietly recoding the material into fewer, larger chunks.

031 min

Where working memory comes from

The term arrived as a deliberate replacement for an older one. Through the 1960s the standard account had a short-term store that passively held material on its way to long-term memory. Baddeley and Hitch argued in 1974 that this was the wrong shape, because a passive store cannot explain how people manipulate what they are holding: reversing a list, comparing two options, keeping a running total.

Their alternative was the multicomponent model described above, and its central claim was in the name. The system is for working with information, not for keeping it.

The capacity figure was revised separately and later. George Miller's 1956 paper put the span at about seven items, a figure that spread far beyond its original claim, and which belongs to its own entry in this library. Cowan's reconsideration in 2001 argued that the real limit, once rehearsal and grouping are ruled out, is closer to four. The two numbers are not in conflict so much as measuring different things: one counts what people can report under ordinary conditions, the other counts what attention can hold when recoding is blocked.

041 min

What working memory changes

The practical figure to carry is three to four chunks, and it is much smaller than most interfaces and documents assume.

Three consequences follow.

Comparison is expensive. Comparing two things requires holding both. Comparing three options on four attributes is twelve items, which is beyond the limit by a wide margin. This is why people comparing plans on a pricing page do not actually compare them; they narrow to two, then to one attribute, and decide there.

Anything not visible is being held. A step in a flow that requires remembering something from the previous step is spending capacity. A form that validates on submit makes the user hold every entry they were unsure about.

Expertise changes the contents, not the size. A senior engineer reading unfamiliar code is subject to the same limit as anyone else. What they have is better chunks, built from years of seeing the same patterns. This is why a new team member is slow in a codebase they are individually capable of understanding.

052 min

How working memory shows up in product, design, and AI

Interface design. Keep the material a decision needs on screen at the moment of the decision. A comparison table beats a carousel, not because tables are prettier, but because a carousel forces the user to hold the first option while reading the second. The same argument decides confirmation dialogs: a dialog that asks "are you sure?" without restating what will happen makes the user retrieve it.

Writing and documentation. A sentence with three nested clauses is not a style problem. The reader has to hold the opening while parsing the middle. This is the mechanism behind most advice about shorter sentences, and it is why the advice is worth following even when the long version is grammatical.

Code review and incident response. Both are working-memory tasks in disguise. A reviewer holds the intended behaviour while reading the actual behaviour. A responder holds a hypothesis while reading logs that may refute it. The reliable interventions are external: write the hypothesis down, keep the diff small, put the timeline in a shared document while the incident is running rather than after.

Prompting and agent design. The parallel is worth naming precisely, because the concepts are related but not identical. A model's context window is a fixed budget of tokens, and a person's working memory is a few chunks refreshed by attention. What transfers is the design advice, not the mechanism: both degrade when asked to hold many items and relate them, and both do better when the material is pre-grouped. Structuring a long prompt into labelled sections helps a reader and a model for different underlying reasons.

062 min

How to work with the limit

Working memory is a capacity rather than a bias, so the useful moves are about spending it well rather than guarding against it.

Build better chunks, do not try to add slots. Capacity in chunks is close to fixed. What is not fixed is what counts as a chunk, and that is what training changes. A naming convention, a well-chosen abstraction, and a shared vocabulary all work by turning several items into one.

Put it outside the head. The cheapest intervention available is writing things down where everyone can see them. A shared document during a meeting, a checklist during a deployment, a whiteboard during a design discussion. None of this is a productivity ritual. It is capacity that no longer has to be spent on holding.

Remove the second task rather than training the first. Load is additive. Someone who must notice, hold and decide at the same time will do all three worse than someone doing them in sequence.

Order matters more than volume. Material presented in the order it will be used costs less than the same material presented in a different order, because nothing has to be held across the gap.

Do not buy the training. Programmes that promise to expand working memory produce reliable gains on the trained task and little else. That claim is the subject of the misunderstandings section below, and it is the single most commercially exploited error about this concept.

072 min

Common misunderstandings

"Working memory training makes you smarter." This is the expensive one. Monica Melby-Lervåg and Charles Hulme reviewed the evidence in Developmental Psychology in 2013, pooling 23 studies with 30 group comparisons. Training produced reliable short-term improvements on working memory tasks themselves. For verbal working memory those gains were not sustained at follow-up. Beyond the trained tasks, they found no convincing evidence of generalisation to nonverbal ability, verbal ability, attention, word decoding or arithmetic. Their conclusion was that the programmes produce short-term, specific training effects that do not generalise.

"The number is seven." Seven plus or minus two is Miller's 1956 estimate and it has its own entry here. The figure people repeat is usually applied to situations Miller was not describing. When recoding into chunks is prevented, the measured limit is closer to four.

"Working memory is short-term memory." They are not the same claim. Short-term memory describes passive holding. Working memory describes holding plus manipulation, and the manipulation is the reason the concept was introduced.

"Capacity is a fixed personal trait like height." Measured capacity varies with sleep, stress, and how well the material is chunked. What is stable is the approximate ceiling, not performance on a given day.

Where the limit is doing useful work. A tight capacity forces abstraction. Chunking is how expertise is stored efficiently, and a system that could hold everything would have less pressure to find the structure that makes holding unnecessary.

082 min

Working memory vs. nearby concepts

The question that separates working memory from cognitive load in practice: are you asking how much a person can hold, or how much this screen is asking them to hold? The first is fixed. The second is yours to change.

092 min

Where the evidence is contested

The capacity number that this entry has been quoting is itself disputed, and the dispute is about the shape of the limit rather than its size.

Wei Ji Ma, Masud Husain and Paul Bays set out the case in Nature Neuroscience in 2014. The traditional view, which they call the slot model, treats working memory as a fixed number of discrete slots — Miller's seven or Cowan's four — where an item is either stored accurately or not stored at all. Their objection is that this does not match what happens when recall precision is measured continuously rather than scored right or wrong.

Stated at its strongest, their argument is that working memory behaves like a limited resource spread across whatever is being held, so that the quality of each representation falls as more items are added. They report that recall variability increases gradually and continuously with set size, which a fixed number of all-or-nothing slots does not predict. On that account, asking how many items fit is the wrong question, because nothing is ever fully in or fully out.

The defence of slot models is that a great deal of behaviour, across many task types, is described well by a small integer, and that the resource account has more free parameters to fit the same data.

Where this leaves a practitioner is more settled than the theory. Both camps agree that performance degrades sharply as items are added, and that the degradation starts at a small number. Whether the fourth item pushes out the first or simply blurs all four, the design response is the same: show fewer things at once, and make each one carry more meaning.

?8 questions

Questions people ask

§7 sources

Sources