MCP tools
Installing registers an MCP server that gives Claude six tools. You do not call these yourself. They are what Claude reaches for when a question needs more than the recall already injected above your prompt.
The three layers
Memory search is deliberately staged, so a broad question does not drag full bodies into context before anything has been chosen.
- Layer 1.
searchreturns a compact index: id, kind, date, title, and one line of the matching body. - Layer 2. The snippet is enough to tell two similarly-titled rows apart and decide which ids are worth reading.
- Layer 3.
get_observationsfetches full bodies, and only for the ids that were picked.
search
Natural language search over this project's memory: past decisions, bugs and reasoning. Takes a
query, an optional kind filter (decision, pattern, bugfix, context, deadend,
preference), a tag, and a project path. Pass "*" to search every project.
get_observations
Full bodies for specific ids returned by search. Takes up to 25 ids at once and a chars budget
per body, so a batch stays readable.
find_usages
The code graph, in the four modes described in Code graph. Takes a symbol, a
mode, and a target for path queries.
Why this beats grep for symbols
It tags the likely definition and re-derives the answer from current source on every call, so it is never stale. Plain grep still wins for anything that is not a code identifier: log files, comments, strings and prose.
remember
Records something outright rather than waiting for it to be captured from a turn. The first line becomes the title. This is what fires when you tell Claude a standing rule, so it survives the session.
"remember that we always use pnpm in this repo"timeline
What happened recently, in order, rather than by relevance. This is what answers "what was I working on" when the prompt itself has no words worth matching.
forget
Deletes observations by id. Used when something was captured that turned out to be wrong, so it stops coming back as context.