Source
Minsky (1974), “A Framework for Representing Knowledge”, MIT AI Memo 306, LIT-410, read in full in NOTE-337 from an OCR of the scan.
NOTE-337 says that no THEORY “cites or should cite this memo as evidence.
It is a source of hypotheses, not of support.” This account is filed
Proposed under the owner’s ruling for that case. It states only what the
reading itself supports: that the monotonicity point is “a correct
observation about classical logic” (NOTE-337, C7), and that the rest is
assertion.
What was actually shown
Defaults (§1.11). A frame’s terminals are never stored empty. They carry “weakly bound default assignments” that new evidence displaces. The memo argues for this from introspection and worked examples (the cube, the room, the kite story), not from a program or an experiment (C1, weak).
The appendix (§6). “Logistic” systems, which separate propositions from general laws of inference, fail for commonsense on five counts: relevancy; monotonicity, “each added axiom means more theorems, none can disappear”; procedure-controlling knowledge, illustrated by the transitivity of “near”; combinatorial explosion; and consistency, which “is not necessary or even desirable in a developing intelligent system”. Its one engaged target is McCarthy’s Airport problem.
What holds. In a monotonic consequence relation an added premise never removes a conclusion. A default displaced by later evidence is a conclusion removed by an added premise. So such a logic, as it stands, cannot represent displacement. That step is definitional, which is why NOTE-337 rates the observation correct.
What is claimed beyond it. That no repair is possible: “I think such attempts will continue to fail” (C7, moderate overall, the no-repair half asserted). That consistency is undesirable, argued by analogy with human inconsistency and the mathematician who “shall not take that step” (C8, assertion).
What this does not say
- That commonsense knowledge is made of frames. The frame account is the memo’s hypothesis, with nothing built or measured, by its author’s account (“Apology!”, p. 2).
- That logic cannot be repaired. Non-monotonic logics were developed to do exactly this. The record holds no reading of them, so it does not say whether they succeed. That claim is THEORY-1239, Proposed.
- How strongly a default is bound. The memo never says how much “strain” a match can bear or when a frame is given up, and it argues against numerical scores (§5.5; NOTE-337, Limitations).
- That consistency is undesirable. That is the memo’s assertion.
Connections
- Context-shifted defaults, measured. Aerts and Gabora’s “pet” experiment (LIT-340) measures the context-shift of exemplar typicality that §1.11 builds into displaceable defaults. The memo predicts only the direction (NOTE-337, Bearing on the record; the pairing is that note’s).
- Family resemblance. THEORY-057 cites the memo’s similarity network (C5) for family resemblance without a definition.
- Hand-coded representations. Chalmers, French and Hofstadter (THEORY-541) name frames among the formats whose filling is left unexplained.
- The later programme. Plain Talk (LIT-399), K-lines (LIT-402) and the jokes memo (LIT-408) build on the frame and its defaults. Gärdenfors’s conceptual spaces (LIT-334, Deferred) are the geometric alternative.