The Magical Number Seven: How Much You Can Hold at Once

Miller's paper is about two separate limits that happen to land on similar numbers. The first is the span of absolute judgment: how many steps of one dimension, such as the pitch of a tone or the strength of a salt solution, a person can label correctly without comparing them side by side. Across experiments that span sat in the neighborhood of seven categories, and adding more steps did not help because people started mislabeling them. The second is the span of immediate memory: how many items can be repeated back straight after they are presented. That too came out near seven.

The important part of the paper is what separates the two. For absolute judgment the limit is in the information: a listener who can only place a tone in one of about seven pitch classes gains nothing from a finer scale. For immediate memory the limit is in the items, not the information they carry. Miller reported Hayes's data showing the span was about nine for binary digits and dropped only to about five for monosyllabic words, even though a word carries many times the information of a one or a zero. The span, as he put it, is almost independent of the number of bits per chunk.

That is what makes recoding so powerful. If memory holds a number of chunks rather than a number of bits, then grouping items into larger familiar units raises how much gets through without raising the count. Nine spoken digits are hard; the same nine digits as three familiar three-digit groups are easy. Chase and Simon showed the same thing in chess in 1973: masters reproduced real game positions far better than beginners did, but with pieces scattered at random the advantage mostly disappeared, because what the master was storing was a handful of familiar patterns rather than a board of separate pieces.

The number itself has not held up. Nelson Cowan reviewed the procedures that stop people from rehearsing items or quietly grouping them, which is what the classic span tasks allow, and concluded in 2001 that the real capacity limit is three to five chunks, averaging about four. Both figures describe the same system; they differ in how much help the task gives. A span of seven is what you get when rehearsal and chunking are available, which is most of the time, and about four is what is left when they are blocked.

George Miller published "The magical number seven, plus or minus two: Some limits on our capacity for processing information" in Psychological Review in 1956, after giving it as an address to the Eastern Psychological Association the year before. It is a review rather than a single experiment, and it opens on a joke: "My problem is that I have been persecuted by an integer." The integer kept turning up in other people's data, and the paper is Miller's attempt to work out whether the recurrences were one finding or several.

The first half gathers absolute judgment studies: tones varying in pitch, tastes varying in salt concentration, points on a line, positions in an interval. The number of categories a person could use reliably clustered in the neighborhood of seven, and Miller proposed the span of absolute judgment as the name for that limit. He then showed two ways around it. Judgments on several dimensions at once carry more information than judgments on one, and splitting a continuum into successive decisions sidesteps the limit altogether.

The second half turns to immediate memory and makes the distinction the paper is remembered for. Memory span is counted in items, so the way to hold more material is to pack more into each item. Miller called this recoding, gave examples from his own experiments on grouping binary digits into larger units, and called recoding an extremely powerful weapon for increasing the amount of information we can deal with. He was explicit that the two sevens might be unrelated, and ended by calling the coincidence of numbers pernicious and Pythagorean.

Nelson Cowan's 2001 review in Behavioral and Brain Sciences is the standard revision. He collected the procedures in which items cannot be rehearsed or silently combined, including tasks that overload the list, block rehearsal with a second job, or use displays too brief to group, and found that capacity settled at three to five chunks with an average near four. The paper did not overturn Miller's argument about chunks; it narrowed how many of them fit.

The rows below are everyday places where a short capacity limit shows, not findings from the studies cited here.

The limit is a fact about the system rather than an error to correct, so the useful moves are about working within it.

• Chunk on purpose. Break a long string into groups of three or four, and give each group a meaning if you can. The count that matters is groups, not characters.
• Write it down the moment it arrives. Anything written is off the span entirely, which is why pilots, nurses and chefs read back and record rather than remember.
• Say numbers in groups when you give them to someone else, and pause between groups so the listener can rehearse.
• Do not hand over more than a few items at once in speech. If there are nine things, give three, confirm, then give the next three.
• Treat seven as an upper bound rather than a target. With rehearsal blocked, which is what a noisy ward or a busy kitchen does, the working figure is closer to four.
• Build familiarity when the material repeats. A chunk is only a chunk once the pattern is known, which is why practice raises how much you can hold in your own field and nowhere else.

Several memory entries describe limits on what comes back, and they are about different parts of the problem.

• It is not the serial-position effect, which is about where in a list an item sat rather than how many items fit. The two show up together: a list longer than the span is where the middle starts to disappear.
• It is not the list-length effect, the finding that any one item is recognized or recalled less well when it was learned in a longer list. That is a statement about the odds for a single item; the span is a statement about the total.
• It is not the Google effect, the tendency to remember where information is stored rather than the information itself. Offloading is a response to a limited span, not the limit itself.