A graph, not a guess.

Most memory tools either match keywords, which misses anything phrased differently, or embed everything with a model, which needs an API key, a download, and trust in a black box. Mneme takes a third path: it builds a graph of the entities in your notes — names, terms, numbers — and lets your question walk it.

The result is a memory that runs entirely on your machine with zero dependencies, answers in milliseconds, and can explain every answer: each result names the entities that carried it. The retrieval idea comes from HippoRAG 2, here over a co-occurrence graph instead of LLM-extracted triples.

pip install git+https://github.com/saranshahuja/mneme

mneme add "Northwind won't approve anything over 50k without a second signature. Budget owner is Dan Whitfield."
mneme add "Dan Whitfield is out the first week of October."

mneme ask "anything blocking us in October?"

1. Dan Whitfield is out the first week of October.
   score 1.00 · matched October
2. Northwind won't approve anything over 50k without a second signature. Budget owner is Dan Whitfield.
   score 0.58 · matched Dan Whitfield

That second result shares no word with the question except a stopword. It surfaces because Dan Whitfield links the two notes — which is the point of the graph. Everything lands in ~/.mneme/memory.db; point elsewhere with --db PATH or MNEME_DB=PATH.

The same four lines from Python: from mneme import Memory, then Memory().add("…") and .ask("…").

Every note you add is scanned for entities with plain heuristics — capitalised spans, content words, numbers and dates. No model, no download, about a millisecond. Entities that appear in the same note get linked, and a link grows stronger each time the pair shows up together.

A question is scanned the same way, and its entities become the starting points of a Personalized PageRank walk over the graph. Notes rank by how much of the walk lands on their entities — so a note two hops away, sharing no words with your question, can still rank second because a person or a project bridges them.

Results are capped, not padded: -k is a maximum, and results scoring far below the top hit are dropped (tunable with --min-score, where 0 restores exact top-k). If nothing in the question is in the graph yet, recall falls back to SQLite full-text search and the result is labeled text match.

mneme add

Remember text, a file (split on blank lines; short blocks are skipped and counted), or stdin via -.

mneme ask

Recall, best first. -k caps results, --min-score sets the relevance floor, --json for machines.

mneme stats

What is remembered: passages, entities, edges, and where the file lives.

mneme forget

Delete one passage by id — its entities and edges are cleaned up with it.

mneme mcp

Serve memory to an agent over MCP. Needs the [mcp] extra.

Every command accepts --db PATH before or after the subcommand, and respects MNEME_DB. Errors are one line on stderr, exit code 1 — pointing it at a PDF tells you so instead of printing a traceback.

pip install 'mneme[mcp] @ git+https://github.com/saranshahuja/mneme'
claude mcp add mneme -- mneme mcp

Cursor and other MCP clients: add {"command": "mneme", "args": ["mcp"]} to your MCP config. The server is read-only — it exposes ask and stats, nothing that writes. You add memories from the terminal, and the agent sees them on its next call.

Everything lives in one SQLite file on your disk. There is no server, no account, no telemetry, and no network code anywhere in the package — the base install has zero dependencies, so there is nothing to phone home with.

Delete a memory with mneme forget and its entities and links go with it. Move your memory by copying the file. The MCP server never sends your filesystem paths to the model it serves.