Consolidation is the part a flat file cannot imitate: it turns a pile of session notes into semantic knowledge, surfaces contradictions, and builds the entity graph on its own. engram_audit is the sharp edge of that — point it at your CLAUDE.md and it tells you which lines the vault now contradicts.
A local memory server: memories go into SQLite with vector embeddings and typed graph edges, and come back by meaning rather than by grep.
- engram_remember stores a memory and auto-extracts the entities and topics in it
- engram_recall and engram_ask — semantic retrieval, or a synthesized answer with confidence and sources
- engram_briefing opens a session with key facts, pending commitments and recent activity
- engram_consolidate distills episodes into semantic knowledge, discovers entities and finds contradictions
- engram_surface and engram_alerts push relevant memories and stale follow-ups without being asked
- engram_audit cross-references an external file such as CLAUDE.md against the vault and flags outdated claims
- engram_connect, engram_forget, engram_entities and engram_stats manage the graph itself
- engram_import_obsidian and engram_import_claude_code bring an existing vault or CLAUDE.md in
- A REST API on http://127.0.0.1:3800 for non-Node environments
A Gemini API key — GEMINI_API_KEY is required, and embeddings are the only network calls the server makes. Install the engram-sdk package globally and run engram init, which registers the MCP server and writes the session hooks; engram doctor checks the vault, the key, embeddings and registration afterwards. Data stays in SQLite under ~/.engram.
One command plus a key — npm install -g engram-sdk, then supply credentials