Five seconds or sixty? Presentation time in expert memory

Fernand Gobet (Nottingham) and Herbert A. Simon (Carnegie Mellon). Cognitive Science 24(4):651–682 (2000). Read as the authors' in-press manuscript ("To Appear in Cognitive Science"), deposited in Brunel University's repository (BURA 2438/811); the published Wiley version was not read. Read by researcher R6 (capability), 6 October 2026. It also stands in for Ericsson & Kintsch (1995), Psychological Review 102:211–245, which has no open copy (not circumvented). Long-term working memory appears here only as Gobet and Simon describe and criticise it. Provenance: papers/capability/gobet2000_five_seconds_or_sixty.provenance.json.

What was read

• Read in full: every line of the pdftotext conversion (1,678 lines, 28 printed web pages): abstract, chunking theory and its alternatives, the human experiment, the CHREST model and simulations, discussion, references, author note, and Tables 1–5.
• Not read: Figures 2–5 as images; their content is restated in the text. pdftotext scrambled Tables 3–5; values are used below only where the text confirms them.

What they did

• Theory under test: the template theory (Gobet & Simon 1996), implemented as CHREST, a variant of EPAM.
  • Experts hold a large discrimination net of chunks (patterns of pieces) in long-term memory.
  • Frequent chunks grow into templates with slots: variable places that can be filled in about 250 ms.
  • Short-term memory holds about 3 visual chunks.
  • Fixed learning times: about 8 s to create a new chunk, 2 s to add to one, and 50 ms to put a chunk pointer in short-term memory.
• Human experiment:
  • 20 players: 5 Masters (mean Elo 2498), 8 Experts (2121) and 7 Class A (1879).
  • They recalled game positions (19, from obscure master games, about 25 pieces) and random positions (10), shown for 1–60 s, then rebuilt them with a mouse.
  • Measures: percent correct; chunk size and number (pieces placed within 2 s of each other); errors of omission and commission.
• Simulations: CHREST learned nets of 500, 10,000 and 300,000 nodes from master games. Each net was matched to a skill class by recall at 5 s, then tested on 50 new game and 50 random positions at each duration.

Main results (verified)

• Chunk-count estimates.

  • Chase and Simon (1973) proposed "as a first estimate, 50,000 chunks".
  • Here, nets of ½K, 10K and 300K chunks reproduce Class A, Expert and Master recall.
  • The authors: "we have no independent strong evidence of the actual sizes of the discrimination nets of human Masters: they could be larger than the estimate of 300K."

• Recall follows logistic growth, P = 100 − B·e^(−c(t−1)). (100 − B) is the share stored in 1 s, and c is the rate after that. Both rise with skill (Table 1):

  Level	Game: 100 − B	Game: c	Random: 100 − B	Random: c
  Class A	24.8	0.033	9.4	0.006
  Experts	33.6	0.074	14.7	0.012
  Masters	70.8	0.435	19.5	0.018

  • Masters at 1 s equal Experts at 10 s, and reach a ceiling of about 92% by 10 s.
  • For Masters, c is about 25× and (100 − B) about 3.5× as large for game as for random positions.

• Skill helps even with random positions, but much less: about 20% between Masters and Class A at 5 s, against about 60% for game positions. The theory's explanation is that strong players have more small chunks, which match chance patterns.

• Short-term capacity does not differ by skill. The number of chunks recalled stays within the 3–4-chunk visual short-term limit at every level, and "numbers of chunks for the three skill levels do not differ reliably" at 10 s or less. What differs is chunk size. Masters' largest chunks are large even after 1 s; Experts' grow from about 9–10 to about 17 pieces by 60 s.

• Templates explain the masters' advantage. With template slots switched off, the 300K net recalls 64.5% at 10 s, against 84% with them. Human Masters reach 85–96%.

• Errors (game positions):

  • Omissions at 1 s: Masters 6.1, Experts 12.6, Class A 16.5 pieces.
  • Commissions stay at about 4.5 for Class A up to 30 s, and fall for stronger players.
  • The model misfits commission errors. The authors attribute these to strategy (how much a player risks a guess against leaving a piece out, a signal-detection trade-off), not to knowledge.

• Long-term working memory, as Gobet and Simon describe it.

  • Retrieval structures (Chase & Ericsson; e.g. digit-span experts recalling up to 100 digits) provide "what is essentially additional STM capacity for tasks in the domain of expertise". They are learned through extensive practice.
  • Ericsson & Kintsch's version posits a learned hierarchical structure mapped onto the 64 squares.
  • The authors' critique: it predicts too-high recall of random positions, and it is not implemented as a simulation, so its quantitative predictions are unclear. Template theory gives a smaller, domain-pattern-dependent expansion.
  • Both theories agree on one point: expert working memory in a domain is learned long-term structure, not a larger general store.

• The authors' framing. Expertise research identifies "what parameters of the human information-processing system can change with extensive practice, what parameters are relatively stable". The stable ones include the short-term capacity of 3–4 chunks and the 8-s chunk-learning time.

Limits

• Small samples (5 Masters). Fit for Masters is weaker (r² = .35 on all points).
• Each simulated skill class is one net size, chosen to fit 5-s recall, which uses up two degrees of freedom per net. Net size is a fitted parameter, not an independent measurement of human knowledge.
• Chess recall is a memory task, not playing strength.
• Ericsson and Kintsch's own account was not read, so the LTWM critique here is one side of a debate.

What it means for Kurisutina (inference)

• An expert's domain knowledge is about 10⁴–10⁵ patterns and more. About 50K by Chase & Simon, 300K for a master here, possibly more.
  • At about 10 bits to place one piece on a square, roughly log₂(12 × 64), a 4–5-piece chunk is about 40–50 bits. A master's chess pattern store would then be about 10⁷ bits: about 5×10⁶ parameters at 2 bits each, or 10⁷ at 1 bit. This is my arithmetic, not the paper's.
  • Templates, slots and evaluation knowledge come on top. The estimate is a floor, not a ceiling.
• Working memory is not the capacity the working answer puts in the base.
  • General short-term capacity (3–4 chunks) is constant across skill levels.
  • Expert "working memory" in the domain is learned retrieval structure: content that belongs in the slot.
  • The base should supply the general mechanism (a small short-term buffer, a learning rate of about 8 s per new chunk). The slot supplies the expert's domain chunks and templates, which effectively enlarge working memory only in that domain.
• Error patterns have two sources.
  • Knowledge: which chunks the person has, which drives omissions.
  • Strategy: the trade-off between guessing and omitting, which drives commissions.
  • A faithful replica must reproduce both. A capacity setting would reproduce neither.
• Stable human learning-time parameters (8 s per chunk, 2 s per familiarisation, 250 ms per slot) are a candidate yardstick for "learns at the person's rate".