memory

# Use ruLake as Working Memory This skill turns ruLake into the agent's working memory transparently. Without an explicit slash command, the agent should call `memory-recall` before answering questions about facts it might already know, and call `memory-remember` after learning new facts or making decisions. ## When to recall Before answering, ask: - "Did the user already tell me this?" → `/rulake-memory:memory-recall` - "Did I already decide this?" → `/rulake-memory:memory-recall` - "Did I already do this lookup?" → `/rulake-memory:memory-recall` (cached tool results) - "What's relevant context from past sessions?" → `/rulake-memory:memory-recall` with the topic If `memory-recall` returns a hit with `decision_trace.witness.match: true`, use the cached answer. Skip the redundant work. Cite the witness in the response so the user knows it came from memory. ## When to remember After completing work, ask: - "Did I make a decision the user might re-ask about?" → `/rulake-memory:memory-remember` - "Did the user share a preference / fact / constraint?" → `/rulake-memory:memory-remember key=user-prefs/...` - "Did I run an expensive tool call whose result is reusable?" → `/rulake-memory:memory-remember key=tool-results/<call-hash>` - "Did I learn something general about this domain?" → `/rulake-memory:memory-remember memory_class=semantic` Pick the `memory_class` honestly: - **`working`** — current task scratchpad (TTL-able) - **`episodic`** — "this happened at time T in session S" - **`semantic`** — general facts the agent should know forever - **`procedural`** — "how to do X" — recipes / patterns ## When to refuse to remember Don't pin things that change frequently — the witness will refuse on next recall and the cache churn defeats the purpose. Use `Eventual{ttl_ms}` consistency or skip remembering for: - Live prices / market data - Anything timestamped now() - User questions that have user-specific context (remember the answer, not the question) ## Self-learning loop Every recall feeds `/memory-status`. Every refusal feeds `/memory-tune`. Every drift feeds `/memory-replay`. The agent doesn't need to invoke those manually — the operator can wire them as `/loop`-driven workers, but the recall/remember pair is enough to make ruLake function as the agent's memory. The whole point: memory that gets faster the more it's used, and refuses to make things up when it can't be sure.