Chase identifies two temporal modes for context updates: hot-path (the agent updates memory while executing its core task — the user prompts it or the harness instructs it to save learnings in real-time) and offline (batch processing recent traces to extract insights — what OpenClaw calls “dreaming”). A third dimension cuts across both: how explicit the update is — whether the user actively prompts the agent to remember, or the agent remembers based on standing instructions in the harness.
This frames your own learning architecture precisely. Your /learn skill is hot-path + explicit (you prompt the agent to extract insights during the session). Your auto-memory system in CLAUDE.md is hot-path + implicit (the harness instructs the agent to save memories without you asking). The /wrap skill is closer to offline — it batch-processes the session’s learnings at the end. OpenClaw’s “dreaming” is fully offline + implicit. This connects to Session capture turns ephemeral AI conversations into a compounding knowledge base — shadcn’s /done pattern is hot-path + explicit, while Evolving summaries beat append-only memory — rewrite profiles, don't accumulate facts describes the storage strategy that offline learning needs (rewrite profiles, don’t accumulate raw facts). The tradeoff: hot-path learning is immediate but adds latency to the primary task; offline learning is thorough but introduces a delay before the agent benefits from what it learned.