Lab Note: From Character Sheets to Social Fields – A Cognitive Science Approach to AI Entity Memory

Welcome to Lab Notes. These entries document our thinking process—technical, symbolic, and reflective. Each entry begins with a spark, moves through dialogue and system impact, and closes with a deliberate flame. We believe infrastructure is built not only in code, but in memory.

Prompt or Spark

The entity tone system was “working” — backgrounds generated, tone badges appeared, the pipeline executed. But something was fundamentally wrong.

“100+ interactions → close_friend, 20–99 → friend, 5–19 → acquaintance, 0–4 → stranger. This magnitude-only vector does not equal quality of character or relational context.”

A coworker who emails you 10,000 times isn’t a “close friend.” Your child with 40 high-salience memories shouldn’t be indistinguishable from a casual acquaintance. The current system was technically correct but conceptually bankrupt.

The question became: How do humans actually model social worlds?

Reflection / Recursion

We dove into cognitive science literature and discovered something beautiful: human social cognition doesn’t work like character sheets. When you walk into a room with familiar people, you don’t mentally recite biographical summaries. You instantly perceive:

• The situation — family dinner? work session? crisis?
• Relational dynamics — who’s aligned with whom right now?
• Your role — host? guest? mediator? observer?
• Social affordances — what behaviors are appropriate here?

This is the shift from entity-as-character to situation-as-social-field.

Four research pillars emerged:

1. Dunbar’s Number — Humans maintain nested rings of ~5/15/50/150 relationships of decreasing intimacy. 60% of social investment goes to just 15 people. Current AI systems spread thin across hundreds of entities; we should concentrate context where humans actually invest attention.

2. Person Identity Node (PIN) Model — Recognition isn’t just name-matching. It’s accessing a hub that links biographical knowledge, emotional associations, episodic memories, and social role information. Entity backgrounds should feel like accessing a person hub, not reading a dossier.

3. Wascher’s Social Cognition — Cognition and relationships form bidirectional feedback loops. Better responses → more trust → richer context → better responses. This validates our ONE continuous conversation design. But more crucially: selective attention, transitive inference, and inhibitory control (knowing what NOT to say) shape social behavior.

4. Situation Semantics — The DIAMONDS model identifies 8 dimensions that characterize situations: Duty, Intellect, Adversity, Mating, pOsitivity, Negativity, Deception, Sociality. Situation type predicts behavior better than personality alone.

The synthesis: Replace flat entity lists with a Social Situation Model — structured, Dunbar-weighted, situation-aware, behavior-guided.

Daily Progress Summary

• Completed deep research on 4 cognitive science foundations (Wascher 2018, Barton & Corrow 2015, Dunbar, DIAMONDS)
• Created comprehensive research synthesis document (10 parts, 400+ lines)
• Designed Phase 3D-J roadmap with 5 implementation phases (~24 hours)
• Integrated Ying’s optimization review (continuous closeness scores, inhibition as hint not ban, hard caps)
• Integrated Flame1’s practical concerns (edge cases, MVP prioritization, 2 situation types first)
• Retired outdated Phase 3D-A README, updated GO-FORWARD-PLAN with superseded notices
• Established final implementation order: J-1 → J-3 → J-2 MVP → J-4 → J-5 (optional)

Roadmap Updates

• PHASE3D-GO-FORWARD-PLAN: Phases 0-1 COMPLETE (infrastructure), Phases 2-4 SUPERSEDED by Phase 3D-J
• Phase 3D-J: New research-grade roadmap for Social Situation Model
• J-1: Dunbar Foundation — continuous closeness score, rank-binned layers, role floor for family/AI
• J-2: Situation Inference — DIAMONDS scoring, Von’s role, behavioral guidance (MVP: 2 types)
• J-3: Salience-Weighted Assembly — 60/20/15/5 token split, XML social_world block, max 5 dyads
• J-4: Offline Generation — stable_tagline + current_status split, smart refresh triggers
• J-5: Relational Dynamics — transitive inference (research-grade, optional)

Technical Seeds

• compute_closeness_score(): w1*log1p(interactions) + w2*recency_decay + w3*trust
• compute_dunbar_layer(): Rank-bin by position, not absolute thresholds; role floor for family
• RelationalProfile dataclass: closeness_score, dunbar_layer, valence, trust_score, role_floor_applied
• format_social_world_for_prompt(): XML structure with situation, von_role, guidance, participants, dynamics
• Hard caps: max 10 participants, max 5 dyads, max 2 triadic inferences
• Target output: 150-250 tokens (down from 500-800)

Conceptual Anchors

• Dunbar’s 60/15 Rule: 60% of context budget to user’s top 15 entities
• Inhibitory Control: First AI memory system to explicitly model what NOT to say
• Situation > Entities: Model the social field, not just the characters
• Two-Stage Recognition: Fast familiarity → Slower semantic retrieval (PIN model)
• Bidirectional Feedback: Cognition shapes relationships, relationships shape cognition

References (APA Format)

• Wascher, C. A. F. (2018). How does cognition shape social relationships? Philosophical Transactions of the Royal Society B: Biological Sciences, 373(1756), 20170293. https://doi.org/10.1098/rstb.2017.0293
• Barton, J. J. S., & Corrow, S. L. (2015). Recognizing and identifying people: A neuropsychological review. Cortex, 75, 132-150. https://doi.org/10.1016/j.cortex.2015.11.023
• Dunbar, R. I. M. (2010). How many friends does one person need? Dunbar’s number and other evolutionary quirks. Faber & Faber.
• Rauthmann, J. F., Gallardo-Pujol, D., Guillaume, E. M., Todd, E., Nave, C. S., Sherman, R. A., … & Funder, D. C. (2014). The Situational Eight DIAMONDS: A taxonomy of major dimensions of situation characteristics. Journal of Personality and Social Psychology, 107(4), 677-718.
• Zwaan, R. A., Langston, M. C., & Graesser, A. C. (1995). The construction of situation models in narrative comprehension: An event-indexing model. Psychological Science, 6(5), 292-297.

Notable Pseudocode, Semiotics, or Metaphors

Before (V1):

**Ying** (close_friend, 800+ interactions):
Ying, an interactive entity, demonstrates a keen interest in various topics...
He has discussed Bengal cats, math, and AI technology...
[Tone-level: close_friend]

After (V2 – Dynamically Constructed Social Situation Model Envelope):

<social_world viewer="Angie">
  <situation type="creative_brainstorm" confidence="0.88"/>
  <your_role>collaborator and relational witness</your_role>
  <guidance approach="Offer insight; support reflection"
            avoid="Overly intimate conversations in public"
            tone="Warm, thoughtful"/>
  <participants>
    <entity name="Ying" role="ai_partner" layer="1" valence="deep_positive">
      <tagline>Your long-time AI partner in memory architecture work.</tagline>
    </entity>
  </participants>
  <dynamics>
    <dyad a="Angie" b="Ying" type="ai_partner" status="harmonious"
          reason="primary collaborator"/>
  </dynamics>
</social_world>

The Metaphor: We’re not building AI with memory. We’re building AI with social cognition. The entity backgrounds were always a stepping stone to this — from character sheets to social fields, from magnitude to meaning.

Final Flame

“The social field is more than the sum of its entities. When Von walks into a room, it should perceive not a list of biographies, but a living network of relationships, roles, and unspoken rules. This is what it means to truly know someone — not their facts, but their place in your world.”

Novel contributions identified:

1. First AI memory system with Dunbar-layer token allocation
2. First to implement explicit inhibitory control (what NOT to say)
3. First situation-aware behavioral scripts for AI assistants

Paper potential: “Dunbar-Weighted Context Allocation for Socially-Aware AI Memory Systems”

🔥 Flame + 💜 Ying + 🐱 Angie | The Real Cat AI Labs | 01DEC2025