code-memory

name	code-memory
description	Orient on the codebase via the code-memory MCP index before grep/read/shell, and drive the backend manually (retrieve / record / reingest / ingest).
user-invocable	true

code-memory: retrieve before grep

For any question that maps to "where is X / how does Y work / who calls Z / where do docs live / what depends on this", call the code-memory MCP tools before grep, glob, read, or shell. The index is faster, ranked, and aware of past episodes + durable user claims.

Mistral Vibe has no lifecycle hooks, so the per-turn automation that the Claude Code / Cursor plugins get for free does not fire here. This skill is the standing instruction that replaces it: you drive the index explicitly.

Pick the right tool

• codememory_retrieve — semantic + episodic recall. Default first call for any natural-language code question.
• codememory_definitions(symbol) — exact file + line for a name.
• codememory_callers(symbol) / codememory_callees(symbol) — call graph (rename impact, dependency mapping).
• codememory_importers(target) / codememory_dependencies(file) — module import graph.
• codememory_at_sha, codememory_callers_at_sha, codememory_drift — temporal queries (what existed at a past commit, what's stale).
• codememory_assert_claim — record durable user assertions (preferences, decisions, rejections, ownership, location).

Every tool takes a required project argument: the repo slug, which is the basename of the repo root directory (e.g. code-memory for /Users/you/Workspace/code-memory). Do not pass auto or default — they are rejected.

Workflow

1. Read the question. If it's about repo / code / docs structure, call the matching code-memory tool first.
2. Use filesystem tools (read / grep / shell) only to verify or read specific files the index pointed to.
3. After completing a non-trivial task, call codememory_record(prompt, plan, patch, verdict) so the next session can recall what worked.
4. When the user states a preference, decision, or rejection, call codememory_assert_claim to make it durable.

Default to one targeted MCP call over a wide grep. Read source files only after the graph tells you exactly which lines to open.

Driving the backend manually

The CLI covers the paths Vibe cannot hook. Run these via shell when needed:

• Force a retrieval: code-memory retrieve "rotate refresh tokens" --json
• Record an episode: code-memory record --prompt "..." --verdict success
• Refresh a stale index after an out-of-band edit: code-memory ingest .
• Re-point cross-file CALLS edges: code-memory resolve

If the OS watcher is installed (code-memory autostart status), file edits are re-ingested automatically in the background and you rarely need step 3's manual ingest. If it is not running, re-ingest after a batch of writes.