The Society of Mind

Marvin Minsky · 1986 · full text · MIT OCW lectures
See also: The Emotion Machine (2006), the sequel that replaces agents with layered critics and "ways to think."

Intelligence emerges from vast numbers of mindless agents organized into agencies. No single agent is intelligent. No single principle explains the whole. The power comes from diversity.

The argument

Minsky's book is 270 one-page essays organized into 30 chapters. The running example throughout is a child building with blocks — a task that looks simple but requires dozens of cooperating processes. Visual parsing, spatial reasoning, motor control, goal management, error recovery.

The core claim: there is no central executive. The mind is a society of agents, each mindless, that compose into agencies. Agencies handle domains (locomotion, language, spatial reasoning) and compete for control through winner-take-all switching. No compromise.

Architectural primitives

Agents and agencies

An agent does one thing. It has no understanding of what it does or why. An agency is a group of agents that together achieve a function none could achieve alone. The mind is agencies all the way up. There is no top.

Proto-specialists are innate agencies that handle survival domains: locomotion, feeding, defense. They are the boot-loader. Everything else is built on top through learning.

Memory agents (the nemes)

Mechanism	What it does	Parts bin analogue
K-line	Records which agents were active during a success. Reactivates that constellation on a similar cue.	Episodic memory / case-based reasoning
Polyneme	A permanent K-line. Activates partial states across many agencies simultaneously.	Embedding vector / pub-sub broadcast
Microneme	A diffuse contextual signal. Sets mood or context globally rather than activating specific representations.	Context injection / system prompt
Pronome	Short-term memory pointer. Controls which role-slot is currently active.	Register / pointer / pronoun
Isonome	Signals many agencies to perform the same operation in parallel.	Broadcast instruction / SIMD
Paranome	Links pronomes across representational realms. Enables cross-domain analogy.	Cross-modal attention / transfer learning

Representational structures

Structure	What it does
Frame	A skeletal data structure with slots and default assumptions. Defaults hold until contradicted by evidence.
Frame-array	Multiple frames for the same object from different viewpoints. Shared slots enable perspective-switching.
Trans-frame	Represents an event: origin, destination, cause, motive, instrument, time. The verb of internal language.
Story-frame	Chains of trans-frames. Narrative as data structure.

Control and error correction

Mechanism	What it does
A-brain / B-brain	A-brain handles world interaction. B-brain monitors A-brain for errors — infinite loops, faulty reasoning. B-brain can veto.
Censor	Suppresses the activity preceding an unproductive action. Preemptive negative expertise.
Suppressor	Inhibits the unproductive action itself. Reactive negative expertise.
Difference-engine	Identifies gap between current state and goal. Invokes an operator known to shrink the gap.
Cross-exclusion	Winner-take-all switching between proto-specialists. One agency suppresses all others. Emotions are cascade effects of this switching.

Key principles

Principle	Statement
Papert's Principle	Mental growth depends not on acquiring new skills, but on acquiring new ways to use existing knowledge.
Exception Principle	Easier to extend old frames by adding exceptions than to replace them.
Recursion Principle	The same strategies apply at every level of decomposition.
Predestined Learning	Sufficient innate constraint makes certain learning outcomes nearly inevitable.

Chapter map

Ch	Title	Key introductions
1	Prologue	Agents, agencies, blocks-world
2–5	Wholes, Conflict, Self, Individuality	Heterarchies, no compromise, circular causality
6	Insight and Introspection	B-brains, self-knowledge is dangerous
7	Problems and Goals	Difference-engines, local responsibility
8	A Theory of Memory	K-lines, level-bands, societies of memories
10–12	Papert's Principle → Learning Meaning	Uniframes, the exception principle
16	Emotion	Proto-specialists, cross-exclusion cascades
18–20	Reasoning → Context	Polynemes, micronemes, connection lines
21–22	Trans-Frames → Expression	Pronomes, trans-frames, isonomes
24–26	Frames → Language-Frames	Frames, defaults, frame-arrays, story-frames
27	Censors and Jokes	Suppressors, censors, negative expertise
29–30	Realms of Thought → Mental Models	Paranomes, cross-realm analogy, recursive self-models

The Emotion Machine (2006)

Twenty years later, Minsky replaced agents with critics. Society of Mind asks how mindless agents compose into intelligence. The Emotion Machine asks how the system decides which agents to run. The answer: layered critics, each monitoring the level below.

Level	What it does	Who has it
1. Instinctive reactions	Hard-wired responses. Pain, reflex, startle.	Every agent
2. Learned reactions	Conditioned responses. Flinch before the blow lands.	Soar (chunking), LLMs (weights)
3. Deliberative thinking	Search over alternatives. Means-ends analysis.	Soar (impasse), ReAct
4. Reflective thinking	Monitor your own deliberation. "Am I stuck?"	Reflexion, RLHF
5. Self-reflective thinking	Reason about your own reflective strategies.	Nobody
6. Self-conscious thinking	Model yourself as others see you.	Nobody

Each level can interrupt, override, or redirect the one beneath it. Emotions are what happens when a critic triggers a large-scale mode switch, shutting down some resources and activating others. Anger narrows attention. Grief suspends goals. Architectural events, not decorations on top of cognition.

The critic hierarchy is A-brain/B-brain scaled to six layers. Soar's impasse mechanism gets two levels (substate monitors the superstate). MAKER gets two (verifier monitors worker). RLHF gets two (reward model monitors policy). Nobody has built six. Minsky argued you need all six before the system can reason about its own reasoning about its own reasoning. Whether he was right depends on whether self-reflection decomposes into a finite stack or requires something else entirely.

Where these ideas landed

Minsky never wrote pseudocode. His contribution is design patterns. But every architecture in this collection reinvented his ideas under different names.

K-line replay is episodic memory. Soar stores a working-memory snapshot at every decision cycle, retrieves by partial cue matching. LLM agents do the same with natural-language episode streams (Generative Agents) and MemGPT's paged retrieval. Record what was active when something worked; replay it when the cue matches.

The A-brain acts; the B-brain watches for errors and vetoes. That's RLHF: a reward model judging the policy model's output. It's MAKER: verifier agents challenging worker outputs. Constitutional AI is a B-brain that reads a constitution. Minsky had the pattern in 1986.

Censors and suppressors are negative expertise: learned knowledge about what not to do. Soar's reject preferences fire before ranking. MemAgent learns what to forget. Alignment filters are suppressors. The censor is more interesting. It suppresses the thought preceding the bad action. Preemptive negative expertise. No modern system does this cleanly.

Cross-exclusion is softmax. One agency wins, all others suppressed. Minsky used it for emotions: proto-specialists competing for control of the whole system. Attention heads compete the same way, but between tokens rather than entire agencies.

Polyneme broadcast is the embedding vector. One signal activates partial states across many agencies at once. Multi-head attention does this with learned projections into subspaces. Minsky's version is simpler and more honest about what it is: pub-sub.

Neighbors

by june.kim Vaswani 2017 · 2 of 5 →