7±2 Is a Myth (Sort Of): What Miller Really Said About Working Memory

Miller’s 1956 paper does not contain the rule people quote from it. It reports two separate limits that both happen to land near seven, and Miller closes by calling that overlap a pernicious Pythagorean coincidence.¹ The limit he did defend is counted in chunks rather than items, and half a century of controlled work put the working figure at about four.²

So the design rule is not seven. Build every dashboard, slide, agenda and onboarding flow for four units, then spend the rest of your effort making each unit carry more.

That single swap changes what you ship this week.

Why a budget, a cliff and a draw

The Wire Model routes a decision to a small ensemble by scoring its features first. Here is what scored:

• Fixed budget on the receiving side (0.9). Your reader’s capacity is not yours to expand. You only choose how many fronts you contest.
• Non-linear failure (0.85). Performance holds flat, then collapses. It does not taper.
• Heterogeneous receivers (0.7). Individual capacity varies enough that two partners in the same meeting keep different subsets.
• Regime-break risk (0.6). Machines now read your document before a human does, and machines have no chunk limit.
• Low strategic content (0.2). Nobody in this problem is bluffing. Signaling and cheap talk have nothing to add, which is why they are absent.

That routes to a resource-allocation game (the budget), a capacity threshold (the cliff), and independent-draw aggregation (the room). Equilibrium, complex and random: three outcome types, so their errors do not lean the same way.

Two members that usually earn a card are folded here. The behavioral layer is the subject of the whole piece rather than a fourth lens, so it adds no lever of its own. Governance shows up inside the portfolio, as the named owner of the cut, because a rule about item counts dies without one.

The framework: a fixed budget, a cliff, and a different four in every head

1. The budget: your reader arrives with four slots and cannot buy more

Start with what Miller actually measured. He ran two separate lines of work. One was absolute judgment, where people label tones, brightnesses or positions, and the average ceiling across the variables he reviewed is about 2.6 bits. The other was immediate memory, where people repeat back a list. His conclusion separates them cleanly: absolute judgment is limited by the amount of information, immediate memory is limited by the number of items.¹

Different limits, different units, near-identical numbers. Miller thought that was an accident and said so in the last line of the paper.

Cowan’s reconsideration treats the seven as a rhetorical device and rebuilds the estimate under conditions that block rehearsal and grouping. Group averages land between three and five chunks, and about four is the central figure.² Prevent covert rehearsal specifically and the number holds near three, whether the units are single words or word pairs the subject learned earlier.⁵

Now treat it as an allocation game. Your reader holds four units of a resource they cannot mint. You choose how many fronts to contest. Contest nine and five of them lose by default, without a fight and without your input.

Miller measured the consolation prize too. Adding attributes to a display raises total capacity while lowering accuracy on any single attribute, so people end up making crude judgments about several things at once.¹ Nine metrics on one screen deliver four facts and a feeling. You do not get to choose the four.

There is a legitimate way to buy more. Visual working memory holds about four objects, but each object can carry several bound features, so sixteen features survive when they are distributed across four objects.⁷ Attributes attached to an object ride free. Attributes listed separately each take a slot.

That is a build instruction. Four cards, each carrying a headline number, its direction and its owner, costs four slots. Twelve tiles cost twelve and deliver four. Same data, different survival rate.

Operators here already know the shape of this. A USSD menu that runs past four options gets abandoned mid-flow. An agent who has to hold float, commission, tariff and settlement date in his head while serving a queue drops one, and it is usually the one you needed. Build the menu, the agent script and the WhatsApp order confirmation to four lines each, and the drop rate moves before you change anything else.

Resource-allocation game, the budget lens

Assumes: the receiver’s capacity is fixed and independent of how badly you want their attention.

Fits because: fixed budget on the receiving side scored 0.9, and the empirical budget is about four chunks.

Breaks when: the reader can pause, re-read and take notes. External memory relaxes the constraint, and a diligence analyst has all three.

Counteracts: the belief that adding a slide adds understanding.

May reinforce: hiding material facts under a four-item cap, which reads as evasion the moment it is found.

2. The cliff: item five is not remembered weakly

The failure is not gradual, which is why intuition gets it wrong. Recall accuracy runs flat across list lengths up to three or four items, then breaks. In the enumeration results Cowan reviews, errors sit near zero for arrays of one to four items and then climb roughly fifteen percentage points per extra item until they reach nearly one hundred percent at eleven.²

The shape of the loss matters even more than the slope. When change detection is modelled properly, the data fit an all-or-none pattern: an item is either held or it is gone, with no faded intermediate copy.⁶ There is no partial credit. Your fifth number is not remembered vaguely. It is not remembered.

Two design consequences follow immediately, and both are free.

First, integrate before delivery. The practical rule from the same literature is to avoid presenting more than a few ideas at once unless they can be integrated fast, and to avoid making people hold unintegrated pieces for long.⁴ A sentence that withholds its unifying theme until the end forces the reader to store three loose fragments and hope. Move the theme to the front and the fragments arrive pre-attached. That is why the conclusion goes first in a memo. Not style. Capacity.

Second, stop splitting attention. Instructional design work has spent forty years on this: separating a diagram from the text that explains it forces mental integration, and physically merging them removes the cost. Presenting the same content twice in two formats adds load rather than insurance.⁸

Translate that into artifacts you own. The legend belongs on the line, not in a box beside it. The denominator belongs on the chart, not in a footnote. The number you want quoted back belongs in the sentence, not in the appendix. Reading a slide aloud while the audience reads it costs you and buys nothing.

Miller reported the same effect from earlier work by Chapman. When observers were told in advance which attribute to report, they were more accurate. Told afterwards, they had to hold all three attributes and accuracy fell.¹ Tell the room what to watch for before the chart appears.

Capacity threshold, the collapse lens

Assumes: performance is flat to a boundary and falls sharply past it, with items held whole or lost whole.

Fits because: non-linear failure scored 0.85, and the error curve past four is steep and measured.

Breaks when: the material is already familiar to the receiver, in which case their chunks are bigger and the boundary sits somewhere you cannot see.

Counteracts: the assumption that a busy screen degrades gracefully.

May reinforce: cutting real signal to hit a number, which is how a dashboard loses the metric that mattered.

3. The draw: each head in the room keeps a different four

Capacity is a population average, not a personal guarantee. Group means cluster between three and five chunks, while individual means spread from two to six.² The partner sitting at two and the partner sitting at six walk out of the same meeting holding different material.

Chunk length compresses the count further. Where the units are longer, such as idioms or short sentences rather than words, adults hold three or four of them.³ Slots also leak: capacity spent holding what is irrelevant to the task is capacity that never reaches your argument.³ The partner rehearsing the question they want to ask is running at reduced width for the next ninety seconds.

Run the arithmetic on the model and the group problem gets sharp. Present nine points to three decision-makers who each retain four at random, and the expected number of points all three retain is nine multiplied by four-ninths cubed. That is under one. The room does not average its memory. It intersects it. Cut to four and the intersection is four, because there was nothing else to sample.

Chunk size is the one variable that genuinely scales, and it is built from the receiver’s prior knowledge rather than from their intelligence. Children who are chess experts out-remember adult novices on board positions.⁴ The cost of a piece of information depends on what the person processing it already knows.⁸

The ceiling on that trick is worth knowing before you rely on it. Subjects trained to repeat strings of eighty or more digits did it by building a hierarchy, and the hierarchy obeyed the same limit at every level: groups of three or four digits, never a mnemonic covering more than five, and never more than four groups inside a supergroup.² Even the record holder builds four of four. Miller’s own demonstration ran the same way, recoding binary digits into larger named units until forty of them came back in a row.¹

So the investor across the table has small chunks in your market and there is no fixing that in one meeting. “Agent churn in the Kisumu corridor” is one chunk to you and five to them. Name it, define it once, then use the identical phrase every time. A synonym is a new chunk and costs a fresh slot.

Independent-draw aggregation, the room lens

Assumes: receivers vary in capacity and sample your material close to independently.

Fits because: heterogeneous receivers scored 0.7, with individual means ranging from two to six chunks.

Breaks when: the group takes shared notes or one dominant voice sets the recap, which correlates the draws and shrinks the variance.

Counteracts: designing for one imagined average reader.

May reinforce: over-repetition, which spends the slots you were trying to protect.

GEER: what to change first, and what each change costs

Four channels carry the load: item count, chunk size, integration and phrasing stability. Start with the free reversible moves.

1. Count, then cut to four. Open your main screen or key slide and count the separately parseable units. Cut to four. Hits item count. Costs an hour.
2. Lead with the conclusion. Put the unifying claim before its supporting parts everywhere, including subject lines. Hits integration. Costs nothing.
3. Freeze the wording. Pick the exact phrase for each of the four and repeat it verbatim in the deck, the update and the dashboard. Hits phrasing stability.
4. Bind attributes to objects. One card carrying number, direction and owner beats three tiles. Hits chunk size.
5. Move labels onto the data. Legends, units and denominators go on the chart. Hits integration.
6. Delete the second copy. Stop reading slides aloud, stop repeating the chart in the paragraph beside it. Hits load.
7. Teach one compound. Give a recurring bundle a name and use it until the receiver treats it as one unit. Hits chunk size. Costs weeks of repetition.
8. Rebuild the reporting layer so four is what exists. Hits item count structurally. Costs a quarter and is the hardest to reverse.

No-lever flag: if the decision genuinely turns on five interacting variables, presentation will not rescue it. Split the decision, or pre-compute the tradeoff into one defensible number and put your name on that number.

RADAR: the redesign portfolio, dated

DO NOW, by T+3 days. Reversible, and they pay off under every scenario.

1. Count the units on your primary dashboard and your key slide. Write both numbers down before you edit anything.
2. Choose the four that decide the business this quarter. One named owner each.
3. Rewrite every top line as the conclusion, with the support underneath it.
4. Run the recall test. Give someone outside the company your four in sixty seconds, talk about something else for a minute, then ask what they remember. What comes back is your real chunk set.

HEDGE, by T+14. Cheap protection against the day someone reads it without you in the room.

1. Lock the exact wording of the four and push it into the deck, the investor update and the reporting tool on the same day.
2. Build two layers instead of one long list: four headlines, each opening to four supporting items.
3. Put units, denominators and periods on the visuals themselves.

DEFER AND TRIGGER. Expensive and hard to reverse, so pre-commit the observable trigger now.

1. Defer: rebuilding the analytics stack, commissioning a full deck redesign, hiring a data visualisation contractor.
2. Trigger to spend: run the recall test on three people. If fewer than three of your four survive with two of them, the structure is at fault. Book the rebuild that week.
3. Counter-trigger: all four come back intact and the decision still stalls. The format is fine and the argument is weak. Stop editing slides and go back to the evidence.

Reading rather than writing. Investors and operating partners sit on the other side of the same constraint. By T+3, write your own four before the founder presents, so their sequencing does not choose your slots for you. By T+14, score memos on a fixed four-field rubric, which turns the draw into a record. Defer changing your screening model until twenty memos have been scored the same way, because before then you are reading your own variance.

CHAIN: the base rate on cutting things down

Pick the comparison by mechanism, not by industry. The class here is every setting where a receiver with fixed capacity has to act on a message of variable length: air traffic readback, surgical checklists, radio number protocol, exam design, emergency dispatch. Their common history is consistent. Interventions that add explanation fail. Interventions that reduce the item count and pre-integrate what remains hold up, which is why checklists are short and readback is scripted.

Second order: cutting to four moves the argument upstream into which four, and that fight is political rather than cognitive. It belongs to the founder and it cannot be delegated to whoever builds the dashboard. Third order: teams start managing the four they publish, so choose four that resist gaming and re-examine them each quarter.

Before crediting the redesign, take out what the act of choosing did on its own. Some of the improvement is that you finally decided which four things matter, and you would have got that from the decision without touching the layout.

Matrix-break flag. Your document’s first reader is increasingly a model, and models have no four-chunk limit. What reaches the human is a summary compressed by something you do not control. Near term, redundancy is free for the machine and expensive for the person, so the two readers want opposite things. Medium term, the answer is to state your four explicitly and early, in stable wording, so any compressor keeps the four you chose. Write for the machine’s indifference and the human’s ceiling in the same document.

What this ensemble cannot see

Begin with what would break the argument outright.

Stakes. These capacity estimates come from lab tasks with nothing riding on them. A partner who wants your deal will re-read, take notes and call you. Motivation buys external memory, and external memory suspends the constraint.

The receiver’s existing structures. Chunk size is built from prior knowledge you cannot observe. You can only test recall after the fact, which tells you the result and never the cause.

Group memory. A room with shared notes, a recording and one loud voice does not behave like three independent draws. The correlation is real and none of these models price it.

The average is not a person. Four is a population figure that ranges from two to six across individuals. You are designing for a distribution, and the person who decides is one sample from it.

All four gaps leave the same instruction standing. Write your four today, in one line each, in the exact words you intend to keep. Then run the sixty-second recall test on one person before your next meeting, and rebuild whatever fails to come back.

Sources and notes

1. Miller, G. A. “The Magical Number Seven, Plus or Minus Two: Some Limits on our Capacity for Processing Information.” Psychological Review 63, 1956, 81-97. Full text at Classics in the History of Psychology, York University. The distinction between the two spans (“Absolute judgment is limited by the amount of information. Immediate memory is limited by the number of items”), the closing verdict (“a pernicious, Pythagorean coincidence”), the crude-judgments-of-several-things finding, the Chapman advance-instruction result and Sidney Smith’s binary recoding demonstration are all in this text.
2. Cowan, N. “The magical number 4 in short-term memory: A reconsideration of mental storage capacity.” Behavioral and Brain Sciences 24(1), 2001, 87-114. Article record with full abstract, and the full target article as PDF from the same host. Section 3.5 gives group means of about 3 to 5 chunks and individual means of 2 to 6. Section 3.3 carries the flat-then-steep performance function and the enumeration error curve reported from Mandler and Shebo (1982). Section 3.4.1 quotes Ericsson et al. (1980) on the trained digit-span subject’s group and supergroup limits.
3. Cowan, N. “The Magical Mystery Four: How Is Working Memory Capacity Limited, and Why?” Current Directions in Psychological Science 19(1), 2010, 51-57. Open access at PubMed Central. Source for the 3 to 5 central store, the 3 or 4 limit on longer verbal chunks such as idioms and short sentences, and capacity lost to task-irrelevant material.
4. Cowan, N. “Working Memory Underpins Cognitive Development, Learning, and Education.” Educational Psychology Review 26(2), 2014, 197-223. Open access at PubMed Central. Source for the practical rule on presenting few items and avoiding held unintegrated information, the reordered-sentence demonstration, and Chi’s (1978) chess result.
5. Chen, Z., and Cowan, N. “Core verbal working-memory capacity: The limit in words retained without covert articulation.” Quarterly Journal of Experimental Psychology 62(7), 2009, 1420-1429. Open access at PubMed Central. Capacity of about three chunks holds across list lengths and list types once covert rehearsal is prevented.
6. Rouder, J. N., Morey, R. D., Cowan, N., Zwilling, C. E., Morey, C. C., and Pratte, M. S. “An assessment of fixed-capacity models of visual working memory.” Proceedings of the National Academy of Sciences 105(16), 2008, 5975-5979. Open access at PubMed Central. Receiver operating characteristic slope of 1.0 indicates all-or-none representation; capacity estimates asymptote at three to four items.
7. Luck, S. J., and Vogel, E. K. “The capacity of visual working memory for features and conjunctions.” Nature 390, 1997, 279-281. Publisher record. The four-object limit and the sixteen-features-across-four-objects result are stated in the abstract on that page; the full text is behind a paywall.
8. Sweller, J., van Merriënboer, J. J. G., and Paas, F. “Cognitive Architecture and Instructional Design: 20 Years Later.” Educational Psychology Review 31, 2019, 261-292. Full text at Springer. Source for the split-attention effect and its removal by physical integration, the redundancy effect, and element interactivity depending on the knowledge of the person processing the information.