Contribution
The paper takes the society-of-agents view as given, credits it to joint work with Papert (Note 1), and asks one engineering question of it: how could very simple agents in one mind communicate? Its answer is that they do not send messages. Arguments live at fixed locations that the agents share. From that one constraint it derives a developmental story about how agents differentiate, how a young mind handles interruption, and how internal “cases” could seed the cases of natural language.
Key insight
If the parts of a mind are too stupid to parse, they cannot talk; they can only share wires. A shared wire works like a shared convention, where water comes from faucets and mail from mailboxes. Conventions are cheap to keep if every new agent is born as a slight variant of an old one, already attached to the same wires. So the constraint that looks crippling, rigid fixed locations, is what development would produce anyway.
Assumptions
- Agents are very simple. They are “just intelligent enough to accomplish their own specialized purposes”, and lack syntax and shared symbol definitions.
- Brain as parallel machine. The scheme is meant to work without the ordering conventions a serial computer would allow.
- Development by splitting. New agents arise “by splitting off from old ones, with only small changes”, so they inherit their data connections. Radically novel agents are assumed rare in infancy, by analogy with organic evolution.
- Method. The paper favours developmental over performance theory: “Only a good theory of the principles of the mind’s development can yield a manageable theory of how it finally comes to work.” It calls its own speculations a possible “model of scientific irresponsibility”.
- Neurology is offered without evidence. c-lines are white matter and agents are cortex. The “laminar hypothesis” has redundant parallel layers that differentiate as crosstalk is reduced, “with no pretense that there is any solid evidence for it”.
Key results
There are no theorems, data or programs. The results are proposals.
-
Five ways to pass arguments.
- attribute-value pairs;
- an ordered list of values;
- an ordered list of pointers;
- a parsed linear message;
- “Send nothing!”
The paper adopts a variant of (5): fixed short-term-memory locations, “global variables, with all the convenience and dangerous side-effects”.
-
STM is the agents. STM is “an extensive, branching structure, whose parts are not interchangeable”. The classic limited-capacity results are reinterpreted as different groups of agents blocking one another’s external communication.
-
Specificity gradient. A few high-level channels span much of the brain; lower agents are segregated into sub-societies that “communicate within, but not between, those divisions”.
-
Temporary storage. Persistence-memory restores a recent sustained state after a transient disturbance. Transient case-shift moves patterns “upward” between c-line levels. The author distrusts the case-shift: it “seems physiologically unnatural to me” and is offered “more as an exercise than as a strong conjecture”.
-
Minitheories 1–3 for descriptions.
- A distinct c-line per property. Fine for infants, “extravagant” for adults.
- A GETPROP-like noun-agent. Rejected, because it loses “homogeneity of symbol-type”.
- Several noun-cases with graded property structures and a case-shifter that moves the focus of attention into a better-described case.
Minitheory 3 predicts a preferred ordering of cases in early language: when OBJECT shifts into SUBJECT, a “third” case replaces OBJECT by default.
-
Language. Children already use something like syntactic structure internally. The puzzle is why grammatical speech comes so late, and the conjecture is a computational step, “learning to translate between languages”. On “innate vs. acquired”, internal uniformities in infancy may “compel society, in subtle ways, to certain conformities”.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | The mind is a society of simple agents whose interactions produce intellect and affect alike | assertion | stated as the background theory (Note 1); not argued in this paper |
| C2 | Simple agents can communicate without messages by reading and writing fixed shared locations | moderate as a design argument | §“Communication”; the computational point (global variables) is sound, and the brain mapping is unargued |
| C3 | Short-term memory is the extensive, non-interchangeable structure of c-lines, and span limits reflect blocking between agent groups | weak | reinterpretation of “well-known experiments”, none cited or analysed |
| C4 | Agents differentiate by splitting from common ancestors, which explains how they come to share connections | weak | analogy with organic evolution; Notes 5–6 conjecture cross-inhibition and “concept-leaf” splitting |
| C5 | A transient case-shift lets young minds handle subproblems without a recursive stack | weak, by the author’s own account | ”something about it bothers me” |
| C6 | Early cognitive cases precede, and shape, the cases of natural language, and may give a preferred case ordering in early speech | weak, testable | conjecture; the paper proposes looking at early language development and reports no look |
| C7 | Grammatical speech arrives late because of an added computational facility, not because internal structure is missing | assertion | analogy with the theory of computation |
Concepts
- Agent. A very simple specialist. “Each such agent is, by itself, very simple.”
- c-line. A communication channel. Each agent reads and writes a few nearby ones, on two or three adjacent levels.
- Specificity gradient. The gradual decentralisation from a few global channels to many local ones.
- Laminae. Redundant parallel layers that act as one unit early and differentiate later.
- Persistence-memory. The tendency of c-lines to restore a recent sustained state.
- Case-shift. A shift of c-line contents into a “more principal” case.
- Cognitive case. A conceptual focus, such as ORIGIN, DESTINATION or INSTRUMENT, with its own property lines. It is the internal analogue of a linguistic case.
- The Society of Minds. The name, in 1977, of the theory pursued with Papert (Note 1).
Connections
- Frames. The paper builds on the frame memo (LIT-410). It is “in part a sequel” to it (Note 1). Its fixed locations extend the frame memo’s “common terminal”. Its adjacent-level memory addressing revisits the frame memo’s two-way frame matching (“I still don’t understand the issues very well”). Families of agents sharing terminals “would usually constitute a ‘frame-system’”.
- K-lines. K-lines (LIT-402) is Minsky’s next step. It recasts these c-lines as K→P connections and says the earlier scheme was “confusingly bidirectional”.
- Named sources. Tinbergen’s hierarchical instinct model, used for coordination and the “consummatory act”. Winston’s near-miss learning. Newell, Shaw and Simon on reading as matching. Halliday’s Learning How to Mean. Hewitt’s antecedent and consequent reasoning.
- Perceptrons, “on tap, not on top”. Local perceptron-like detectors are proposed for learning symbol patterns. Minsky’s own book limited what simple perceptrons can do, and he answers that limit here: inputs are already meaningful symbols (Note 8).
Bearing on the record
- Collectives and wholes. For the record’s work on collectives, notably LIT-046’s computational account of collective intelligence, this paper supplies the inverse case. It is a collective whose members are mindless and cannot know more than their superiors (Note 9).
- Group minds. The society analogy here does not license treating a group of minded people as a mind, which is the move Schwitzgebel’s argument needs (LIT-159). The pairing is mine.
- Language. C6 bears on the lexicalism and morphology papers in the record (LIT-031, LIT-071) only remotely. It is a claim about acquisition order, not structure, and no THEORY here rests on it.
- Boundary. No machine-learning practice instruction; no anthology topic holds it.
Note of 2026-10-16: THEORY-043 now cites LIT-399 as a source.
Limitations
- No evidence of any kind: no experiment, no program, no citation for the STM claims.
- The neuroanatomy is explicitly unsupported.
- The paper leaves “conflict and control in the Society of Minds” aside. It treats only communication, so the theory it names is not stated here, only presupposed.
- Its strongest mechanism, the case-shift, is the one its author trusts least.
Open questions
- Is there a preferred case ordering in early language development, of the kind C6 predicts? The paper asks this and does not look.
- Do STM span limits behave like blocking between agent groups, differing by context, as C3 needs? Or do they behave like a fixed number of shared slots?
- Can fixed-location communication scale beyond infancy without the case-shift, or some other way to rebind roles?
Corrections
- none to a seeded skim (there was no seed).
- Citation. The brief is right on number, year and venue, but it understates the first appearance. MIT AI Memo 430 is dated June 1977. The IJCAI-77 text appeared in August 1977 (vol. 2, pp. 1083–1092) and is the same text.
- Attribution. “Generally taken as the first published statement” of the society idea is not something the paper claims. It presents the theory as unpublished and in progress, under the name “The Society of Minds”, plural. Minsky’s 1979 K-lines memo credits a principle of the theory to Minsky and Papert’s 1974 Oregon lectures, which were not checked, so priority is left open here.