Marvin Minsky (1980), Cognitive Science 4(2):117–133 — DOI-10.1207/s15516709cog0402_1. First issued as MIT Artificial Intelligence Laboratory, A.I. Memo No. 516, June 1979, 23 pp. — https://hdl.handle.net/1721.1/5739. This filing was read from the memo.
Key takeaways
- Memory re-creates a state. “The function of a memory is to re-create a state of mind.” A K-line is attached to the agents active during a memorable event, and reactivating it re-imposes that partial mental state, a subset of the agents’ states.
- Dispositions before propositions. Feelings, attitudes and “ways of seeing things” may be the simpler elements, and facts the harder ones. So memory is first a predisposition, not a stored description.
- Level-band principle. A K-line should reach only a band of levels below its own. Reaching too low imposes false perceptions; reaching too high makes one “hallucinate the present problem as already solved”. The memo calls this “probably the most important idea of this theory”.
- K-recursion. New K-lines attach mainly to K-nodes already active, so “new memories are composed mainly of ingredients from earlier memories”.
- Abstraction for free. With cross-exclusion groups, conflicting details cancel out, so an “accumulating” K-node extracts what its instances share.
- Learning needs three nets. A goal net G must control how K learns to drive P. Global, recency-based reinforcement cannot solve human credit assignment.
Standing in the record
Filed on 2026-10-01 at the owner’s request, as part of Minsky’s Society of Mind programme. It is the programme’s theory of memory. It is also the first of Minsky’s texts filed here to use the phrase “Society of Mind”, in its section “MENTAL STATES and the SOCIETY of MIND”. Its Note 1 calls the theory one “I have been evolving jointly with S. Papert”, and the acknowledgement says “the basic idea came in conversations with him”.
It builds on two works in this filing, and says so.
- Plain Talk. It extends Plain Talk (LIT-399), “which the present paper complements in several areas”. That paper’s c-lines “correspond roughly to the K—>P connections here”. It recasts their “confusingly bidirectional” structure as a K–P duality.
- Frames. It extends the frame memo (LIT-410) by offering a distributed implementation of frames (Note 9). “A K-node acts like a ‘frame’”: the agents it activates in its level band are the frame’s obligatory terminals, and weak connections at the band’s lower fringe give “the loosely bound ‘default assignments’”.
The jokes memo (LIT-408) takes over its partial mental states. The 1986 book, filed alongside by another pass, makes K-lines one of its central mechanisms.
NOTE-352 is the close reading of 2026-10-01, from the memo, and it placed the work: Active. It is a clear, compact statement of memory as state reinstatement with a principled limit on how much to reinstate. It is not evidence. The constructive part stops halfway: the P→K connection, which relates perceptual events to goals, is left “somehow”.
It belongs with three works already in this record, which the memo does not cite.
- Convergence zones. Damasio’s proposal (LIT-385, Deferred) has recall as the time-locked re-activation of feature fragments in early cortices, directed by convergence zones. That is the same shape as a K-line re-activating agents.
- Somatic markers. The somatic-marker hypothesis (LIT-384) has learned situation–body-state links “re-enacted” to bias choice. That is a disposition-first memory of the kind this memo argues for.
- Sparse codes. For its crossbar problem the memo’s Note 10 proposes sparse random subset codes on a shared bundle of lines, after Mooers and Willshaw et al. That is the family of codes behind sparse distributed memory, which this record reads in LIT-269 and NOTE-242.
All three pairings are mine. The memo carries no instruction for machine-learning practice, and no anthology topic holds it.