Identity Persistence

All agents drift from their persona — we measure it with EWMA SyncScore, re-inject SOUL.md before it's gone.

You wrote a SOUL.md. By turn 30 the agent sounds like every other LLM. CORTEX won't let that happen.

The persona is loaded from ~/.openclaw/SOUL.md at startup and prepended to every prompt at top priority — your agent never wakes up not knowing who it is. Every N turns (default 10) an EWMA SyncScore evaluates how far recent output has drifted; when it falls below threshold (default 0.6), the persona block is re-injected mid-context automatically. Behavioral observations are extracted to a JSONL log at ~/.openclaw/cortex/observations.jsonl so you can see exactly when and why drift happens. Personality nudges from Learned Intuition (if active) fold in before injection.

Install

openclaw plugins install @globalcaos/openclaw-identity-persistence

Drop a persona at ~/.openclaw/SOUL.md — a default is bootstrapped if absent. Then enable in openclaw.json:

"plugins": {
  "allow": ["tinkerclaw-identity-persistence"],
  "entries": { "tinkerclaw-identity-persistence": { "enabled": true } }
}

Pairs Well With

• @globalcaos/openclaw-learned-intuition — failure-derived personality nudges flow into SOUL.md re-injections. Your agent doesn't just dodge bad actions; the patterns reshape who it is.
• @globalcaos/openclaw-computational-humor — persona-aware humor. Your SOUL.md tells LIMBIC what flavor of jokes match your agent, instead of generic LLM banter.
• @globalcaos/openclaw-total-recall — episodic memory shaped by persona. ENGRAM stores the events; CORTEX makes sure recall stays in character.

👉 https://github.com/globalcaos/tinkerclaw 👉 https://thetinkerzone.com

Clone it. Fork it. Break it. Make it yours.