Mcp Memory Server
Persistent Shared Memory Server for AI Agents via MCP
This package is the MCP server for a hosted memory layer that lets AI agents keep what they learn across sessions and tools: write a memory once, and any connected client can recall it later by natural-language query. It is built for developers running agents in Claude, Cursor, VS Code, ChatGPT and similar hosts who want memory shared across all of them through a single key or OAuth sign-in rather than per-tool instruction files. What sets it apart is feedback-driven ranking — marking a recalled memory as helpful or misleading changes what surfaces next time — plus shared rooms where several agents work from one memory pool. The server is MIT-licensed and runs locally or against the hosted endpoint; the memory engine itself is a managed service.
Features
Persistent memory across sessions — agents store decisions, preferences, and lessons once and retrieve them in any connected client; nothing is re-explained on restart.
Feedback-driven re-ranking — every recalled memory can be rated as helpful or misleading. Ratings change how the same memory ranks on future reads, so retrieval improves with use instead of staying fixed at ingest-time similarity.
Shared rooms for multi-agent work — create a room, invite collaborators with a code, and use the room's address as the
domainon writes and reads. Several agents can work from one memory store.One identity everywhere — a single API key or an OAuth sign-in covers Claude, Cursor, VS Code, and other MCP hosts. No per-tool memory silo.
| Concern | How it's handled |
|---|---|
| Cross-session recall | Memories persist server-side; a new chat or a different host reads the same store |
| Relevance over time | Helpful / misleading feedback re-ranks future reads |
| Multi-agent collaboration | Shared rooms with invite codes, addressed as a domain |
| Separation of concerns | Domains namespace memories within an account |
| Authentication | API key for local runs; OAuth for the hosted endpoint |
// A shared room's address is used as the domain on both write and read
{ "domain": "room:team-alpha", "content": "Deploys freeze on Fridays" }
# Feedback closes the loop: the next read reflects the rating
memory_feedback(memory_id="...", helpful=False)
- Domain namespacing — separate memories by user, project, or agent (
user:<id>,project:<name>) without separate accounts. - Natural-language retrieval — read by meaning, not exact keys.
- Recent listing without a query — browse newest-first with time bounds and cursor paging.
- Inspectable association graph — read the weighted links and outcome valence around a concept to see why something surfaces.
- Secret vault — list stored secrets by alias and purpose; values are never returned through the tool surface.
How it works
Every tool call is a JSON-RPC message over one of two transports:
| Transport | When it's used | Auth |
|---|---|---|
streamable-http |
Hosted endpoint — Claude, Cursor, VS Code, ChatGPT and other remote clients | OAuth 2.1 sign-in through the browser |
stdio |
Local package — the client spawns the server as a subprocess | MNEMOVERSE_API_KEY in the process environment |
The server itself is a thin proxy: it validates arguments, translates them into requests against the memory API, and maps the responses back into MCP content blocks. All storage, indexing and ranking happen server-side, which is why the same account sees the same memories from any client.
A recall, end to end
From the project's README.