Working Memory

What it is

The number everybody knows is wrong. "The Magical Number Seven, Plus or Minus Two", published by George A. Miller in Psychological Review 63(2) in 1956, is among the most cited papers psychology has produced, and its headline figure survives in business books, design guidance and pub arguments long after the field moved on. Nelson Cowan's 2001 review in Behavioral and Brain Sciences 24(1), 87-114, puts the working figure at roughly three to four items. That is a large correction to a small number, and almost nobody outside the literature has heard about it.

What the store does matters more than its exact size. Long-term memory is effectively unbounded; nobody has run out of room. Working memory is not, and it sits directly in the path of everything on its way there. Nicholas Carr's image for this is a thimble filling a bathtub. The bath is long-term memory. The thimble is what you can carry across in one trip, and information arriving faster than the thimble can move it is not stored more slowly. It is spilled. Nothing queues.

That bottleneck is the whole basis of cognitive load theory, John Sweller's 1988 account of why presentation changes how much of identical material a person learns. Load theory divides the demand on the channel into the intrinsic difficulty of the material, the extraneous demand imposed by how it is laid out, and the productive effort of building understanding. Extraneous load is the part that can be designed away, and every hyperlink in a paragraph, every unnecessary field in a form and every decision an interface asks for is a small withdrawal from the same account.

The far side of the bottleneck is memory consolidation. Holding something briefly involves temporary chemical changes at synapses that already exist. Making it permanent requires the brain to grow new connections, which takes time and protein synthesis, and Eric Kandel's work in Aplysia counted them. Working memory is the stage before that construction starts, which is why a limit on the thimble is a limit on what can be built at all.

One thing Miller's paper is careful about and its popular version is not: the unit is a chunk rather than an item, and what counts as a single chunk depends on what the person already knows. A chess position is many pieces to a novice and one shape to a master. The capacity limit is real and narrow. What can be packed into each slot is not fixed, and expertise mostly works by packing better rather than by adding slots.

Carr gets the revision right in The Shallows, noting the downward correction rather than repeating the headline, and that is worth crediting in a genre where "seven plus or minus two" is one of the most durable stale figures in circulation. The vault records his chapter 7 as cited accurately.

In effect

Miller's seven, and Cowan's three

The research case is the correction itself. Miller's 1956 paper established that immediate recall breaks down at a small and specific number of chunks, which was the finding that made capacity a measurable property rather than a metaphor. Cowan's 2001 review pulled the number down to three or four by looking at what happens when people are prevented from using rehearsal and grouping strategies to inflate their apparent capacity. Both results are in the literature. Only one of them made it into general circulation, and it was the older, larger and more memorable one, which is a small case study in how a figure survives its own refutation by being easier to quote.

Six digits, and the walk from the phone to the keyboard

The everyday version runs several times a day. A verification code arrives as six digits, and you hold them across the two seconds it takes to switch windows, and sometimes you do not. Someone reads out a phone number while you are looking for a pen. You go to the kitchen for four things and return with three. Nothing in any of that is a memory problem in the clinical sense; it is the ordinary size of the thimble, and interference is the usual cause of the failure rather than decay. This is also why interruption is so expensive, and why attention residue and switch cost are the same limit seen from another angle. Whatever was being held is competing with what has just arrived.

The seven-item menu, and a design rule that outlived its number

Interface guidance absorbed Miller's number decades ago and has been slow to give it back. The advice to keep a navigation menu to about seven items, and its many descendants, is a design heuristic that has been carrying a psychology citation it no longer has any right to. The practice underneath the rule is sound: progressive disclosure, sensible defaults, splitting a long form into steps that each hold one idea, and stripping optional fields from a checkout are all ways of spending a narrow channel on the decision the user actually came to make. Erping Zhu's 1999 study in the Journal of Educational Multimedia and Hypermedia 8(3), 331-358, found comprehension falling as link density rose, and readers' confidence in their own comprehension falling with it. That is a single study from 1999, and the vault holds no independent replication of it, so it is the illustration here rather than the evidence. The load argument does not need it. The capacity limit is what carries the weight, and the capacity limit is not in dispute.

What it does not say

It does not say seven. Any page, deck or product spec citing the magical number as the current figure is citing 1956 and ignoring 2001, and the correct number for a design decision is nearer three or four.

It does not say the capacity can be enlarged by practice. This vault has not assessed the working-memory training literature, and no claim about it is made here in either direction. The chunking point is a different claim and a narrower one: what fits into a slot depends on what you already know, which is a statement about knowledge rather than about capacity.

It does not say devices are eating your working memory. The smartphone brain-drain finding, that a phone's mere presence reduces available capacity, largely fails meta-analytically, at d = -0.02 across 166 effect sizes from 53 samples. A separate meta-analysis of 56 studies and 7,093 participants in Media Psychology finds null summary effects across cognitive functions with a single exception, and the exception is working-memory capacity specifically, at an effect far smaller than the original. That survivor is worth stating precisely because it is the only one, and because our note records that meta-analysis without author names, so this site can report its conclusion and cannot yet cite it properly. See cognitive offloading for the rest of that literature and its condition.

It does not resolve what working memory is made of. The capacity limit is one of the better-established results in cognitive psychology and is why this page carries robust. The architecture underneath it, how many stores there are and how attention relates to them, is a live research question that this page does not enter.

Who developed it

George A. Miller, 1920 to 2012, one of the founders of cognitive psychology, published the capacity work in 1956 while at Harvard, and has no page on this site yet. Nelson Cowan, at the University of Missouri, is the source of the revised lower estimate, and has no page here either. John Sweller built cognitive load theory on the limit in 1988, and has no page on this site yet. Nicholas Carr is the writer who carried the material, including the downward revision, to a general readership.

---