memory_read with no arguments returns only the index; a topic's full content comes back when you name it, and memory_search pulls a single fact without dragging the rest in with it. Writing is deferred rather than immediate — memory_append_session stores a summary and the daemon distils it afterwards, which is why the tool asks for durable findings and decisions at 300-800 tokens rather than a play-by-play, since a longer summary costs more at consolidation. If the daemon is not running, summaries keep accumulating and nothing turns them into memories.
A file-backed agent memory: a MEMORY.md index, topic files behind it, and a sessions directory that a consolidation daemon reads.
- memory_read, which loads the MEMORY.md index on its own — the cheap call — and pulls the full content of named topic files only when you ask for them.
- memory_search, a substring search across the memory files for recalling one fact without loading everything.
- memory_append_session, which appends a session summary to the sessions directory for the daemon to extract durable memories from later.
- memory_daemon_status, which reports whether the consolidation daemon is running, not installed, or unavailable on this platform.
MEMORY_DIRECTORY and SESSION_DIRECTORY, the directories the index and the session summaries live in. The consolidation daemon, for summaries to become durable memories — memory_daemon_status is how you find out whether it is there.
One command — npm install -g mcp-agent-memory