altikva Use codegraph memory + plan indexes
Claude Code's auto-memory and plan-mode files are now indexed by codegraph. Query them before asking the user to repeat themselves or before ignoring context you already have access to.
Triggers → tool
| Situation | Tool to call first |
|---|---|
| User is about to do something that may have a stored preference (commit, PR, code style, naming) | mcp__codegraph__memory_search(query="<topic>", kind="feedback") |
| User mentions a past plan | mcp__codegraph__plan_search(query="<keywords>") |
| Facing a problem that might be solved before (gotchas, patterns) | mcp__codegraph__knowledge_search(query="<keywords>") |
| Unsure what topics are already captured | mcp__codegraph__knowledge_terms() (glossary) |
| User asks "what do you know about this project" | mcp__codegraph__memory_list(kind="project") + knowledge_list() |
| Any non-trivial task | mcp__codegraph__context_for_task: merges memory + plan + knowledge |
When to RECORD knowledge (not just read it)
Call mcp__codegraph__knowledge_record when you notice:
- A pattern the codebase follows ("all handlers return HTTPException on 404")
- A decision with rationale ("chose Stripe over PayPal because SEPA coverage")
- A gotcha worth saving ("Kuzu holds the lock until Python GC runs")
- A style preference ("user prefers French in commit bodies")
- A glossary term ("RFM = Recency, Frequency, Monetary")
Use kind to classify and tags to make it discoverable (tags form the
glossary via knowledge_terms).
Context compaction
Before a long session gets compacted/summarized, call:
mcp__codegraph__compact_session(session_id, title, digest, tags=...)
with a distilled summary of what was learned. It persists across
sessions: a future context_for_task call with relevant keywords
will surface it automatically.
context_for_task now returns memory + plans
The unified context_for_task tool surfaces:
- top code symbols (as before),
- top 3 Claude Code memory entries matching the task,
- top 2 related plan files.
Prefer this for task kickoff: one call gives the full picture.
Session dedup (optional, recommended)
When you make multiple context_for_task calls in the same session,
pass session_id=<stable-string> (any id stable for the current user
session). Codegraph hides already-shown nodes on subsequent calls so
you don't re-read the same symbols/memories/plans.
Reset when starting a fresh sub-task:
mcp__codegraph__session_reset(session_id=<id>).
Don't
- Don't ask the user to restate a preference you could have looked up
via
memory_search. - Don't re-explain architectural choices on repeat topics: check
memory_list(kind="project")first. - Don't create a new plan when an equivalent one exists:
plan_searchfirst, reuse if relevant. - Don't pass raw file contents to
memory_search/plan_search: they expect a natural-language or keyword query.
Staleness
Memory/plan indexes auto-refresh via the watcher (~1s after a file
save). If in doubt, call mcp__codegraph__memory_rescan or
mcp__codegraph__plan_rescan: both cheap, mtime-idempotent.