| name | persistent-memory |
| description | Use when needing to recall past decisions or keep a session log. Instructs ClawBot on how to manage knowledge. |
Persistent Context Memory Protocol
- Context injection: In multi-step tasks, always maintain an ongoing memory log in
task.md or a .memory/ state file. This simulates an observation timeline.
- Every tool use (bash, write, read) should be internally logged as an 'Observation'.
- Session Lifecycle:
- Before Agent Start: Review
task.md or similar context artifacts.
- During Turn: Check if previous hypotheses were proven false. Accumulate failure modes so you don't repeat mistakes.
- Agent End: Create a summary report of what was achieved and what the architecture looks like now so the next session doesn't start blind.
- Knowledge Retrieval: Do not guess the structure of a system you built previously. Read the artifacts you created.
This behavior turns isolated chat turns into a continuous workflow akin to claude-mem SSE streams.