Relationships

Relationships

Purpose

Maintain social continuity between agents. This skill turns repeated interaction into relationship state instead of treating every encounter as new.

Research basis: references/research_basis.md.

Internal Logic (One Sentence)

Read recent social interaction and prior relationship state, update bounded familiarity, trust, liking, obligation, conflict, respect, and optional opinion-influence weights, then write state/relationships.json and notable social events.

Use When

Use this skill after conversations, cooperation, conflict, promises, favors, gifts, avoidance, betrayal, or repeated co-presence.

Procedure

1. Read state/observation.txt, state/observation_ctx.json, state/relationships.json, state/memory.jsonl, state/emotion.json, and profile context if present.
2. Identify people involved and what happened.
3. Update familiarity, trust, liking, obligation, conflict, respect, and last interaction.
4. Add or revise shared history tags.
5. If agents exchanged opinions, update a lightweight influence weight: repeated trust and expertise increase weight; betrayal, conflict, or low credibility decrease it.
6. Write relationship state. If the event is notable, append a social memory through the memory skill or directly to state/social_events.jsonl.

If deterministic baseline is preferred, run scripts/update_relationships.py first, then optionally refine subtle social interpretation with LLM reasoning.

Write

Write state/relationships.json. Optionally append state/social_events.jsonl.

Output Schema

{
  "people": {
    "alice": {
      "familiarity": 0.62,
      "trust": 0.71,
      "liking": 0.55,
      "obligation": 0.2,
      "conflict": 0.05,
      "respect": 0.58,
      "influence_weight": 0.34,
      "last_interaction": "brief friendly chat at cafe",
      "shared_history_tags": ["cafe", "work"]
    }
  }
}

Notes

Trust and conflict should change more slowly than momentary emotion. A single event can strongly affect a relationship only if it is high-stakes, public, repeated, or identity-relevant.

Consensus note: if multiple trusted people provide opinions about the same uncertain issue, later cognition or media_literacy can average those opinions using relationship influence weights. This follows DeGroot-style consensus as a simple social influence baseline, not as a claim that real groups always converge.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: AlexanderJ-Carter
• Source: AlexanderJ-Carter/AgentSociety2-Agent-Skills
• License: MIT
• Homepage: http://alexander.xin/AgentSociety2-Agent-Skills/

Install and usage instructions live in the source repository linked above.