OpenClaw Memory v4.5.4 (2026-04-07)

File-based memory system for single-user AI agents: HOT/WARM/COLD/BUFFER + SCRATCH. ~12KB, 9 files, zero dependencies.

Execution

Runtime Decision Tree (for AI execution)

On session start:

1. Read MEMORY.md. Output: HOT context loaded.
2. Read memory/working-buffer.md. Output: current task state loaded.
3. If memory/working-buffer.md.pending exists, recover memory/working-buffer.md.pending before any review or maintenance. Output: pending rotation data recovered or explicit decision that no pending recovery is needed.
4. If memory/working-buffer.md is stale (>2h) and unfinished, recover pending work first. Output: either recovered task state or explicit decision that no recovery is needed.
5. Read episodic notes for today and yesterday. Output: recent events and decisions loaded.
6. Scan the last 20 lines of .learnings/ERRORS.md, .learnings/LEARNINGS.md, and .learnings/FEATURE_REQUESTS.md if those files exist. Output: matching lessons loaded or no-op if nothing relevant exists.
7. Resume from file contents. Stop when the active task or current request has enough memory context to continue safely.