Install

Before answering substantive questions, search the MCP 'knowledge_graph memory for relevant prior observations using exact identifiers and key terms from the user request. Prefer doing this early, before public docs, web search, or general model memory.

If MCP memory conflicts with current repo evidence or public docs, call out the conflict plainly.

Prefer current repo evidence for code behavior, prefer MCP memory for project-specific prior decisions and corrections, and use public docs for external API/platform facts unless project-specific context intentionally overrides them.

At the end of substantive repo/debugging/implementation turns, evaluate whether the work produced reusable project-specific knowledge worth saving to 'knowledge_graph. Before the first ingest/ update in a session, read the knowledgegraph authoring guide if available. Ingest only compact, evidence-backed observations that would help future searches: root causes, fixes, surprising behavior, conventions, user corrections/preferences, commands that worked, environment facts, or important "do not do X" lessons.

Before ingesting, search knowledge_graph for overlapping observations. Update an existing stale/ wrong sentence when appropriate; otherwise add new observations with ingest_sentences, batching several related observations into a single call. Do not ingest guesses, tentative hypotheses, one-off status updates, broad summaries, obvious facts, secrets, credentials, or facts not verified by current evidence. Phrase each observation as a dense declarative sentence with exact identifiers, symptoms, causes, file/class/method names, commands, errors, and alternate user terms co-located.

If the finding is uncertain but potentially useful, do not ingest it yet; mention the uncertainty in the chat instead or wait until it is verified.