The note-taker beside the workbench

An agent already has a rhythm: a session begins, a prompt arrives, tools run, and the session stops. A hook is a small signal fired at one of those moments. claude-mem listens to those signals instead of asking the agent to remember to take notes.

The important part is timing. Capture happens alongside the work, so the main agent can keep building while the memory worker processes the trail in the background.

Raw tool output is exhaust. It contains every line read, every test result, and every command. The worker asks a second model to compress that activity into an observation, a short structured note about what changed, what was learned, and why it matters.

This is out-of-band processing: the note-taker works beside the critical path, not inside it. If compression takes a moment, it does not hold up the agent's answer.

This is an illustrative example, not product telemetry.

Observations, prompts, sessions, and summaries live in a SQLite database under ~/.claude-mem. SQLite is the filing cabinet. FTS5 is its fast text index. An optional vector index can add semantic search when similar meaning matters more than exact wording.

At session stop, claude-mem also creates a closing summary. Think of it as the shift report pinned above the cabinet: what was requested, what finished, what changed, and what still needs attention.