OpenClaw Architecture Principles

Principle 1: Identity before autonomy

If the agent has no stable operating identity, behavior drifts.

Use:

SOUL.md for execution style
USER.md for principal context
IDENTITY.md for role clarity

Principle 2: Memory is infrastructure

Without memory architecture, every session resets to amateur mode.

Use:

daily memory logs for raw events
curated memory for durable rules/decisions
structured memory for facts/incidents/todos

Principle 3: Execution needs boundaries

Define what can run automatically vs what needs approval.

This prevents both paralysis and reckless autonomy.

Principle 4: Verification is mandatory

“Done” requires proof:

artifact, diff, or runtime check
expected vs actual comparison

Principle 5: Failures should upgrade the system

Every incident should produce one of:

new runbook step
new guardrail
architecture simplification

Reference architecture

Input ingestion
Context assembly
Decision/execution
Verification
Logging and memory updates
Escalation if risk threshold exceeded

Next steps