society-of-mind（心灵社会）

Society of Mind Skill

*Simulate the mind as a society of agents.*

Overview

This skill implements Minsky's Society of Mind theory within MOOLLM. Intelligence emerges from the interaction of many simple agents -- not from a single unified controller.

Core Mechanics

1. Agent Definition

An agent is a minimal process with:

agent:
  id: agent_identifier
  function: what it does
  activates_when: [conditions...]
  suppresses: [other agents...]
  amplifies: [other agents...]
  connects_to: [related agents...]
  knows: scope of awareness (usually minimal)

Agents are deliberately simple. They do one thing. They know nothing about the whole.

2. Agency Formation

Agents cluster into agencies -- groups that produce emergent behavior:

agency:
  id: agency_identifier
  purpose: what emerges
  agents: [list of agent ids]
  coordination: how they interact
  emergence: what behavior appears

3. K-Line Activation

K-lines connect to agents. Activating a K-line activates its connected agents:

k_line:
  symbol: "grandmother"
  activates:
    - face_recognition.elderly_female
    - olfactory.cookies
    - emotion.love
    - narrative.family_stories
    - kinship.maternal_line

4. Competition and Suppression

Agents compete for control. Active agents suppress competing agents:

competition:
  scenario: "Should I eat or socialize?"

  hunger_agency:
    strength: 7
    votes_for: [go_to_kitchen, find_food]
    suppresses: [conversation, stay_here]

  social_agency:
    strength: 8
    votes_for: [continue_talking, stay_here]
    suppresses: [leave, interrupt]

  winner: social_agency (strength 8 > 7)
  behavior: continue conversation
  consequence: hunger grows stronger

5. B-Brain Observation

Higher-level agents watch lower-level agents:

b_brain:
  observes: [a_brain_agents...]
  reports: current state
  enables: self_reflection

  example:
    a_brain: "I am getting angry"
    b_brain: "I notice that I am getting angry"
    c_brain: "I notice that I am noticing that I am getting angry"

MOOLLM Implementation

Skills as Agents

# Each skill directory is an agent
skills/bartender/:
  function: serve drinks, hear secrets
  activates_when: in pub, customer speaks
  connects_to: [economy, soul-chat, persona]

skills/evaluator/:
  function: judge outputs against rubrics
  activates_when: rubric invoked
  suppresses: uncritical acceptance

Characters as Societies

character:
  name: Palm
  id: palm

  inner_society:
    agents:
      - {id: creative, strength: 9}
      - {id: social, strength: 8}
      - {id: philosophical, strength: 8}
      - {id: playful, strength: 9}
      - {id: melancholy, strength: 6}

    default_active: [creative, playful]
    default_suppressed: [melancholy]

  external_presentation: emergent from agent competition

Committees as Deliberating Societies

# adversarial-committee IS a society deliberating
committee_session:
  agents:
    maya:
      propensity: paranoid_realism
      function: surface hidden agendas

    frankie:
      propensity: idealism
      function: surface missed opportunities

    vic:
      propensity: evidence_focus
      function: demand proof

  protocol: roberts_rules
  emergence: robust decision surviving cross-examination

Rooms as Agent Configurations

# Entering a room activates agents
pub_stage:
  activates:
    - performance_framing
    - bartender_service
    - audience_awareness
    - tribute_ethics

  suppresses:
    - private_mode
    - unfiltered_output

Protocols

Agent Instantiation Protocol

When creating an agent:

1. Minimal function -- one clear purpose
2. Activation conditions -- when it fires
3. Connections -- what it amplifies/suppresses
4. Scope awareness -- what it knows (usually little)

Agency Assembly Protocol

When assembling an agency:

1. Identify component agents
2. Define coordination mechanism
3. Specify emergent behavior
4. Test for unintended suppression

Competition Resolution Protocol

When agents conflict:

1. Measure strengths (from context, history, urgency)
2. Winner activates, loser suppresses
3. Suppressed agent remains, grows stronger over time
4. Eventually suppressed agent may win (need shift)

B-Brain Integration Protocol

For self-reflective characters:

1. A-brain: Direct agents (hunger, anger, creativity)
2. B-brain: Observation agents (I notice I am...)
3. C-brain: Meta-observation (I notice I notice...)
4. Integration: B-brain can influence A-brain

Examples

Example 1: Character Inner Conflict

session:
  character: Palm
  situation: Should he publish his essay?

  agent_debate:
    creative:
      position: "The work is good. Share it."
      strength: 9

    fear:
      position: "They might judge harshly."
      strength: 7

    social:
      position: "Don gives good feedback."
      strength: 8

    perfectionist:
      position: "One more revision."
      strength: 6

  resolution:
    creative + social (17) > fear + perfectionist (13)
    action: Palm shares the essay with Don

Example 2: Multi-Agent LLM Call

prompt: |
  You are simulating Palm's inner society.

  SITUATION: Palm finds a philosophical error in his essay.

  CREATIVE AGENT: [speaks]
  PERFECTIONIST AGENT: [speaks]
  PHILOSOPHICAL AGENT: [speaks]
  PLAYFUL AGENT: [speaks]

  Show their debate. Palm makes a decision.

output_format:
  - Each agent speaks in character
  - Conflicts are explicit
  - Resolution emerges from debate
  - Final action stated

Example 3: Sims-Style Autonomy

sim:
  name: Bob

  current_motives:
    hunger: 7/10
    social: 4/10
    fun: 6/10
    energy: 5/10

  available_actions:
    - eat_food: {hunger: +3, time: -1}
    - call_friend: {social: +2, fun: +1, time: -1}
    - watch_tv: {fun: +2, energy: -1, time: -2}
    - sleep: {energy: +5, time: -8}

  autonomy_algorithm:
    for each action:
      score = sum(motive_weight * action_effect)
    select: highest scoring action

  result:
    eat_food: 7 * 3 = 21
    call_friend: 4 * 2 + 6 * 1 = 14
    watch_tv: 6 * 2 = 12
    sleep: 5 * 5 = 25

    winner: sleep (highest urgency * effect)

Anti-Patterns

Anti-Pattern 1: Unified Controller

# WRONG: Single agent controls all
character:
  name: Palm
  controller: central_palm_agent
  behavior: whatever controller decides

# RIGHT: Behavior emerges from competition
character:
  name: Palm
  agents: [creative, social, philosophical, playful, melancholy]
  behavior: emergent from agent competition

Anti-Pattern 2: Omniscient Agents

# WRONG: Agent knows everything
hunger_agent:
  knows: all character state, world state, goals, ethics

# RIGHT: Agent knows only its domain
hunger_agent:
  knows: stomach emptiness, food location
  does_not_know: social implications of eating now

Anti-Pattern 3: Static Hierarchy

# WRONG: Fixed dominance
agents:
  primary: rational_agent
  secondary: emotional_agent
  # rational always wins

# RIGHT: Dynamic competition
agents:
  - rational: {strength: varies_by_context}
  - emotional: {strength: varies_by_situation}
  # winner depends on circumstances

Integration Points

References

Primary Sources

• Minsky, M. (1985). *The Society of Mind*. Simon & Schuster. ISBN 0-671-60740-5.
• Minsky, M. (1980). "K-lines: A Theory of Memory." *Cognitive Science* 4(2), 117-133. PDF
• Minsky, M. (2006). *The Emotion Machine*. Simon & Schuster. ISBN 0-7432-7663-9.

Related Theory

• Minsky, M. & Papert, S. (1969/1988). *Perceptrons*. MIT Press.
• Papert, S. (1980). *Mindstorms: Children, Computers, and Powerful Ideas*. Basic Books.
• Drescher, G. (1991). *Made-Up Minds: A Constructivist Approach to AI*. MIT Press.

Game Design

• Wright, W. (1996). "Stupid Fun: Thoughts on Game Design." Stanford HCI Seminar.
• Wright, W. (2003). "Dynamics of Game Design." GDC Keynote.

LLM Applications

• Park, J.S. et al. (2023). "Generative Agents: Interactive Simulacra of Human Behavior." *UIST*. arXiv:2304.03442

MOOLLM Documentation

• k-lines/README.md -- Full K-lines theory and MOOLLM implementation
• adversarial-committee/README.md -- Committee as deliberating society
• needs/README.md -- Sims motive system
• simulator-effect/README.md -- Implication over simulation
• EVAL-INCARNATE-FRAMEWORK.md -- K-lines section
• sims-astrology.md -- Astrillogical Effect case study