Skip to main content

​
Goal

Learn manager and operator patterns over time without unsafe self-modification.

​
Learning Pipeline

  1. Observe:
    • completed tasks
    • approvals/rejections
    • communication edits
  2. Synthesize:
    • summarize patterns into operating profile
  3. Approve:
    • manager reviews profile changes
  4. Apply:
    • update prompts/skill policy only after approval

​
Data Files

  • Input:
    • /docs/agents/learning/interaction-log.jsonl
  • Output:
    • /docs/agents/learning/OPERATING_PROFILE.md

​
Command

pnpm openclaw:learn

​
Safety Guarantees

  1. No autonomous policy mutation.
  2. No secrets in learning data.
  3. External communication privacy rules remain fixed.
  4. Human signoff required before behavior promotion.