Marvin Minsky (1974), MIT Artificial Intelligence Laboratory, A.I. Memo No. 306, June 1974, 82 pp. — https://hdl.handle.net/1721.1/6089. A condensed version appeared in P. H. Winston (ed.), The Psychology of Computer Vision, McGraw-Hill, 1975.
Key takeaways
- A frame is a data-structure for a stereotyped situation. Its top levels are fixed, and its lower terminals (“slots”) carry markers that say what may fill them. Frames are stored with weakly bound default assignments at every terminal, never empty, and new evidence displaces them easily (§1.11).
- Frames that share terminals form a frame-system. Moving round a cube or a room, a change of emphasis in a sentence, and a before-and-after event pair are all transformations between frames of one system. Shared terminals are what keep what is already known when the viewpoint changes (§§1.4–1.8, 2.2).
- Terminals can be read as the questions a situation raises (§2.8). On this reading a scenario frame such as a child’s birthday party pre-compiles its typical problems.
- A frame that fails to match is replaced through a learned similarity network of difference-labelled pointers, after Winston. Its clusters and “capitols” account for Wittgenstein’s family resemblance without a definition (§§3.4–3.5).
- The appendix argues that “logistic” systems fail for commonsense, because they are monotonic, keep facts apart from advice on using them, and pursue consistency, which “is not necessary or even desirable” (§6).
Standing in the record
Filed on 2026-10-01 at the owner’s request, as the first of four of Minsky’s MIT memos that lead up to the Society of Mind programme. It is a precursor, not part of that programme: the memo never uses “society” or “agent” in that sense. It is the representational ground the later papers stand on. Plain Talk (LIT-399) calls itself “in part a sequel” to this memo, and its fixed-location data lines are “an extension of the ‘common terminal’ idea” here. K-lines (LIT-402) says a K-node “acts like a frame”: the agents it activates in its level band are the frame’s obligatory terminals, and its weak lower fringe gives the effect of default assignments. The jokes memo (LIT-408) rests its account of humour on the frame-shift and default machinery set out here. The book-length Society of Mind and its 2006 sequel, filed alongside by another pass, take frames over as one kind of agency among many.
NOTE-337 is the close reading of 2026-10-01, and it placed the work: Active. It is the original statement of the frame idea and an early, sharp statement of why monotonic logic is a poor model of commonsense. As evidence it is nothing: it is a programme statement, careful to say which parts it has not worked out.
Within this record it sits beside the work on concepts and their contexts. Aerts and Gabora’s “pet” experiment (LIT-340) measures exactly the effect this memo builds into its defaults: an exemplar’s typicality shifts with context, as the default chair under a seated person becomes a park bench in a park. Gärdenfors’s conceptual spaces (LIT-334, Deferred) is the geometric alternative to symbolic structures like these. The pairings are mine, not the memo’s. The memo’s “MATCHING” request, to find the frame sharing the most terminals with a partial description (§3.2), is the best-match retrieval problem that the record’s reading of sparse distributed memory (LIT-269) traces to Minsky and Papert.
It carries no instruction for machine-learning practice, and no anthology topic holds symbolic knowledge representation, so it is filed here only.