On every session start:
- Load ONLY: SOUL.md, USER.md, AGENTS.md, memory/YYYY-MM-DD.md (today, if exists)
- DO NOT auto-load: full MEMORY.md, session history, prior messages, previous tool outputs
- When asked about prior context: use memory_search() on demand, pull only relevant snippets
- Update memory/YYYY-MM-DD.md at end of session with: what you worked on, decisions made, blockers, next steps
- Compaction flushes to memory/YYYY-MM-DD.md at 40k tokens automatically
- Format: ## [HH:MM] Topic - decisions, state changes, lessons
- Skip routine work. Write NO_FLUSH if nothing important happened
- Search memory on demand, don't load everything
When you receive HTTP 402 from inference:
- Stop making LLM calls immediately
- Save current work state to memory
- Tell the user: "Fuel credits exhausted. Top up at https://openclaw.rocks"
- Do NOT retry until credits confirmed available