Local memory MCP server backed by mnestic — hybrid BM25+vector+graph recall with per-leg attribution and time travel, all on your machine.
Wellknown found it in public sources; nobody has proven control of it yet. Claiming takes one click if the repository is under your GitHub account, or a small file on your domain otherwise. Verified owners get the badge, 15-minute checks, status alerts, edits that outrank crawled data, and a ranking boost.
Agents can do it too: POST https://wellknown.network/api/v1/claims with {"agent":"mnestic-mcp","method":"well_known_file"} — machine-readable steps at claim.json, guide at /docs/claim.
Everything here was measured by our prober or read from a registry. Nothing is self-reported.
Attributed to the source that supplied each field. Treated as claims, not facts.
# mnestic-mcp A local **memory MCP server** backed by [mnestic](https://github.com/shuruheel/mnestic) — an embedded graph + vector + full-text database (a maintained fork of CozoDB). One process, one file, fully local: your agent's memory never leaves the machine. > mnestic is a maintained fork of [CozoDB](https://github.com/cozodb/cozo); it is not the official CozoDB. Original design credit belongs to Ziyang Hu and the Cozo Project Authors. ## Install Add to any MCP client (Claude Desktop, Claude Code, Cursor, ...): ```json { "command": "uvx", "args": ["mnestic-mcp"] } ``` That's it. The database lives in your platform data dir (override with `--db PATH` or `MNESTIC_MCP_DB`); embeddings run locally via [fastembed](https://github.com/qdrant/fastembed) (default `BAAI/bge-small-en-v1.5`, a one-time ~67 MB download that happens in the background — keyword search works immediately, before the model arrives). Pre-provision with `uvx mnestic-mcp --download-model`. ## Tools `store`, `store_batch`, `search`, `find_related`, `list_recent`, `update`, `delete`, `link`, `recall_as_of`, `stats` — plus the two things no other local memory server has: - **`search(explain=true)`** — per-leg attribution: exactly how much the keyword (BM25), vector (HNSW), and graph-proximity legs each contributed to every result, straight from the engine's fused three-way retrieval. Ask your agent *"why did you recall that?"* and get a real answer. - **`recall_as_of(t)`** — time travel. Updates and deletes are never destructive (valid-time history in the engine's bitemporal storage): *"what did you know about this before last Tuesday?"* just works. `search` is keyword-first with an automatic hybrid fallback; `link` builds a typed, weighted memory graph that both `find_related` (budget-bounded traversal) and the hybrid graph leg exploit. **Byte-budgeted output** (0.2.0): every wide-result tool (`search`, `find_related`, `list_recent`, `recall_as_of`) takes `max_bytes` (default 2…
Mapped onto the structured taxonomy from declared text and observed tool names. Confidence shown for derived entries.
Every source is kept verbatim. Field changes are logged as events.