OpenClaw Configuration Deep Dive
Objective
Turn a working install into a reliable operating system.
The 4 config layers that matter
1) Identity layer
Files:
SOUL.mdUSER.mdIDENTITY.md
Rule: keep voice/mission specific and stable. Avoid generic fluff.
2) Memory layer
Files:
memory/YYYY-MM-DD.md(raw timeline)MEMORY.md(curated durable memory)- optional structured memory (
facts.json,decisions.jsonl,todos.jsonl)
Rule: if it matters later, write it down now.
3) Execution layer
Define:
- what can run autonomously
- what needs approval
- what is always blocked
Rule: high-risk actions require explicit user approval and rollback plan.
4) Verification layer
Every non-trivial task should end with:
- artifact/diff proof
- runtime verification
- concise report of done vs not done
Recommended default config stance
- tight permissions at start
- explicit runbooks for risky tools
- strong “truth in execution” policy (no fake done)
Common anti-patterns
- giant unstructured system prompt
- no memory maintenance discipline
- no post-incident documentation