Reflection

Reflection

Purpose

Convert repeated experiences into higher-level understanding. This skill lets the agent learn patterns about self, others, places, routines, culture, and consequences.

Internal Logic (One Sentence)

Compress repeated memories, emotions, plans, relationship changes, and routine outcomes into concise reflections, then append state/reflections.jsonl and update beliefs, preferences, or self-concept when evidence is strong.

Research basis: references/research_basis.md.

Use When

Use this skill periodically, after major events, after repeated similar memories, at the end of a day, or before long-horizon planning.

Procedure

1. Read state/memory.jsonl, memory/memory.jsonl, state/relationships.json, state/routine.json, state/emotion.json, state/physiology.json, and existing state/reflections.jsonl if present.
2. Look for repeated patterns, contradictions, strong emotions, completed goals, failures, and social consequences.
3. Create concise reflections that can guide future choices.
4. Update beliefs, preferences, and self-concept only when evidence is strong enough.
5. Write new reflections and current belief state.

Write

Append state/reflections.jsonl. Write state/beliefs.json, state/preferences.json, and state/self_concept.json when changed.

Output Schema

Reflection JSONL line:

{
  "tick": 180,
  "type": "pattern",
  "summary": "I often become tired and irritable after skipping lunch.",
  "evidence": ["missed lunch", "anger rose", "plan interrupted"],
  "confidence": 0.74,
  "tags": ["physiology", "routine", "emotion"]
}

state/beliefs.json:

{
  "self": {
    "late_meals_reduce_mood": {
      "belief": "Skipping lunch tends to make me tired and irritable.",
      "confidence": 0.74
    }
  },
  "others": {},
  "places": {}
}

Notes

Do not reflect every tick. Reflection should be slower than memory. It should compress experience into useful patterns without erasing the raw memory trail.

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.