Digit Span, Chunking, and Why "Seven Plus or Minus Two" Is Not a Limit
The famous digit-span figure is about bits, not digits — and the practical limit was later revised downward to roughly four. Understanding that difference explains what a number memory score actually tells you.
Almost every popular account of this cites Miller's 1956 paper and stops. The paper does not say what people think it says.
What Miller actually argued
George Miller's 1956 paper examined span in terms of information content, using the idea that memory capacity is limited by the amount of information it can process. His famous result was that span runs to aboutseven items plus or minus two.
Three qualifications usually get dropped:
- It is not a count of digits. The unit was information — essentially bits. Miller argued the bottleneck sat in the channel rather than in storage, and that techniques such as recoding and mnemonic devices could break through it.
- The "items" were often chunks, not single symbols. Some of his data came from lists of words already grouped into syllables. That is closer to chunking than to reading digits one at a time.
- It was later revised downward. Cowan's 2001 re-analysis argued that capacity is closer to four chunks once you require the items to survive a distractor task — which is the closer analogue of what a digit span test actually asks you to do.
Both papers are cited on our sources page. The short version: "seven" was never a ceiling, and for raw digit recall the better reference is four.
What this test actually scores
The number memory test presents a string of digits for a fixed time, then clears it and asks you to type what you saw. Correct recall increases the length by one digit; any error ends the run. Score is the longest string recalled correctly.
Note what that does and does not include. There is no distractor task, no delay between presentation and recall, and no interleaving. So it sits at the easier end of digit span protocols, which is exactly why scores here run higher than the revised four-chunk figure.
The exposure rule
Longer strings are shown for longer, adding 700 ms per digit up to a 9-second ceiling. The table below is computed from the rule the script actually implements.
| Digits | On screen |
|---|---|
| 3 | 3.6 s |
| 4 | 4.3 s |
| 5 | 5.0 s |
| 6 | 5.7 s |
| 7 | 6.4 s |
| 8 | 7.1 s |
| 9 | 7.8 s |
| 10 | 8.5 s |
| 11 | 9.0 s |
From public/test-scripts/number-memory-test.js. One practical consequence: because the exposure scales, the test partly compensates for slower reading at longer lengths. A 10-digit string is not scored under the same effective pressure as a 3-digit one.
Chunking is the whole game
If capacity is counted in chunks rather than digits, then recalling more digits means using fewer chunks. This is not a trick and it is not cheating — it is the mechanism the original research identified.
| Stored as | Chunks | Verdict |
|---|---|---|
| 1 · 4 · 9 · 2 · 7 · 3 · 8 · 5 · 6 | 9 | Beyond a 4-chunk span — usually fails |
| 1492 · 738 · 56 | 3 | Comfortably within span |
Same nine digits, one third of the storage cost. This is why people who chunk routinely score well past the raw-digit expectation, and why a high score reflects strategy at least as much as capacity.
Reusable groupings
The groupings below cost no time to apply and cover most digit strings. Recognising one is instant; the whole point is that it happens faster than reading individual digits.
- 1.Years. Any four digits in the range 1000–2999 is plausibly a year. Instant chunk.
- 2.Repeated pairs. 1212, 3434, 5656 — mirror pairs read as one object.
- 3.Phone-number grouping. Three digits, then three, then four — the grouping is already known.
- 4.Sequences and doubles. 1234, 4321, 1111, 6789.
- 5.Dates and times. 1400, 0930, 2345 — formats are pre-chunked in memory.
In practice, one reliable grouping is worth more than four unreliable ones. Learning to spot repeated pairs takes a couple of sessions; memorising a mnemonic list takes longer and generalises worse.
What a score does not tell you
- Not general memory. One narrow visual span under one exposure rule.
- Not comparable across rules. A stricter protocol with a distractor gives lower numbers for the same person.
- Not a stable trait on its own. Attention, sleep and familiarity with the test all move it.
- Not an assessment. This is not a diagnostic tool and cannot evaluate anyone’s cognition.
Test it, then try to beat it
Run it once without a strategy to get your raw span. Then try the groupings above and see how far chunking moves it — the gap between those two runs is the part you can actually influence.
Sources: Miller (1956),Cowan (2001). Exposure times are this site’s own rule frompublic/test-scripts/number-memory-test.js. See the sources page for what this site does and does not claim.