Mnemosyne MCP¶
Every memory system in this catalogue can add a fact. Almost none of them can tell you which earlier answers are now wrong because that fact changed.
Server type: Local (stdio)
Install: npx -y -p mnemosy-ai@next mnemosy mcp --db ... --workspace ... --agent ...
Package: mnemosy-ai 2.0.0 stable, 3.0.0-rc.2 on @next, MIT
Runtime: Node 22.16.0 or newer
Store: one SQLite file per workspace
Recall: hybrid retrieval with cross-encoder reranking
Requires: no model API key, no database server, no Docker
Why This Matters for Operators¶
The failure mode of agent memory is not forgetting. It is remembering something that stopped being true and applying it confidently to a decision six weeks later. A price changed, a contact left, a policy was superseded, and the memory keeps serving the old version because nothing in the store knows the two are related.
This server is built around that relationship. It keeps an evidence-backed store, and a correction retires the memories that depended on it rather than sitting beside them as a newer opinion. The vendor also claims exact replay and bisect, which is the property that makes a memory auditable: you can return to a state and take it apart.
For an operator the value is not that an assistant remembers more. It is that when a number changes, the assistant stops quoting the old one, and that a disagreement about what it "knew" can be settled by replaying the state rather than by argument.
What It Does¶
One SQLite file per workspace holds the memories. Recall is hybrid, with cross-encoder reranking on top of retrieval, so a query matches on meaning as well as on tokens.
claude mcp add mnemosyne -- npx -y -p mnemosy-ai@next mnemosy mcp \
--db ./.mnemosyne/memory.sqlite --workspace my-project --agent assistant
The local engine needs no model API key, no database server and no Docker, which matters for a component holding a business's accumulated context: it can run entirely on the machine that already has the data.
What Was Verified¶
The npm registry entry for mnemosy-ai resolved on 29 October 2026 at version 2.0.0, licence MIT, repository github.com/28naem-del/mnemosyne, binary mnemosy. The package description states the design in one line: a local-first long-term memory for AI agents with an evidence-backed SQLite store, hybrid recall, cross-encoder reranking, and corrections that retire stale memories.
The vendor notes that @next is a 3.0.0-rc.2 pre-release and that the stable 2.0.0 exposes the same MCP tools, so a production install can stay on the stable tag.
Limits Worth Recording¶
An early-stage package from a single maintainer, so the surface should be expected to move. The reranking and retirement behaviour were recorded from the published description rather than driven against a live store this pass.
@next in the install line is a pre-release tag. Pin the stable version unless there is a specific reason not to.
Memory scope is a design decision. One file per workspace keeps projects apart, and the workspace and agent flags in the install line are the mechanism, so naming them carelessly is how two projects end up sharing a brain.
FAQ¶
Do I need an API key for an embedding model?¶
No. The local engine needs no model API key, no database server and no Docker, which is what makes it usable on a machine that already holds the data.
What happens when a fact changes?¶
A correction retires the memories that depended on it, rather than adding a newer entry beside the old one, so the stale fact stops being served.
Can I tell why the agent remembered something?¶
Yes. The store is evidence-backed, and the vendor exposes exact replay and bisect, so a state can be returned to and taken apart.
Which version should I install?¶
mnemosy-ai@2.0.0 is the stable release and exposes the same MCP tools as the 3.0.0-rc.2 pre-release on @next.
Where does the data live?¶
In a single SQLite file per workspace, on the machine running the server, passed with the --db flag.
Integration with CorpusIQ¶
CorpusIQ is the connector layer that lets AI assistants read the systems a business already runs on - accounting, payments, CRM, email, files and analytics - and answer with the source named. It is read-only by design: it retrieves data and never writes to, or acts in, a connected system, which is what makes it safe to point at live accounts.
This page is part of the CorpusIQ directory, the reference we maintain for the MCP server and agent-skill ecosystem. Run alongside CorpusIQ, a third-party server like this one can be combined with a company's own business data in a single answer.