The magical number seven is a misquote

What the paper actually said

Miller’s 1956 paper described three separate limits that happen to land near seven, and argued they are different kinds of limitation. The span of absolute judgment — how many levels along one dimension you can tell apart — is limited by information, about 2.5 bits worth. The span of immediate memory is limited by the number of items. The span of attention is a third thing again.

Miller wrote that he suspected the recurring seven was "only a pernicious, Pythagorean coincidence". He was reportedly unhappy about how the paper came to be remembered, and his own emphasis was on chunking — recoding information so each item carries more — rather than on the number.

The modern estimate is closer to four

When you design a task that prevents chunking, rehearsal and support from long-term memory, capacity comes out at about four items, with the commonly quoted range being three to five. The argument is that Miller’s seven included all the spontaneous chunking people do, so it measured performance rather than raw capacity.

Four is not settled either. There is published work arguing the answer depends on what you are counting, and some researchers still defend a figure nearer seven. What everyone agrees on is that the number is task-dependent, which is a more useful thing to know than any single value.

Chunking is how anyone beats the limit

Read "482917" as six digits and it sits at the edge of what you can hold. Read it as "four eighty-two, ninety-one seven" and it is three items. Nothing about your memory changed; the units did.

The classic demonstration is in chess. Masters reconstruct board positions far better than beginners — but on randomly arranged pieces the advantage nearly disappears. It is worth being precise here, because the usual telling overstates it: later work found masters retain a small but real advantage even on random positions. The chunking account survives, slightly bruised.

The training study everyone should know about

An undergraduate known in the literature as SF practised memorising random digit strings for an hour a day, several days a week, for more than a year and a half — over 230 hours of laboratory practice. His digit span went from 7 to 79.

His method was recoding into things he already knew. A competitive runner, he heard "3492" as three minutes and 49.2 seconds, near a world-record mile time.

And here is the part that matters. His span for letters stayed at about six. The gain was entirely specific to digits and transferred to nothing. It is the same conclusion as the research on aim training and on brain training generally: you get very good at the thing you practise, and the improvement stops at its edges.