Mnemoverse Memory

What does this agent do, and when should you use it?

Mnemoverse Memory is an MIT-licensed TypeScript MCP server that connects AI assistants to the hosted Mnemoverse memory engine. It runs as a local stdio server and makes authenticated network calls to core.mnemoverse.com; the server is open source, but the underlying memory engine has no self-hosted build. Its ten MCP tools write, search, and list memories, collect outcome feedback, report statistics, manage shared rooms, and list Vault secret metadata without exposing secret values. Memories tied to the same account can follow a user across connected clients including Claude Code, Cursor, VS Code, and ChatGPT. Helpful or misleading feedback changes future recall ranking through a Rescorla-Wagner prediction-error update rather than relying only on static similarity. It is a good fit for users seeking managed cross-tool continuity, but not for organizations that require memory processing and storage entirely within their own perimeter.

An MCP client launches @mnemoverse/mcp-memory-server, which forwards only explicitly supplied tool-call data to the hosted API. memory_write stores content, concepts, and a domain; memory_read searches in natural language with optional domain, time, author-exclusion, result-count, and ordering controls; memory_list_recent returns a cursor-paginated chronological feed. memory_feedback submits memory IDs, an outcome score, and an optional domain so later recall can be re-ranked. memory_create_room, memory_invite_to_room, memory_join_room, and memory_list_rooms provide explicitly authorized shared domains; memory_stats reports memory and domain information, while vault_list exposes only secret aliases and purposes. Clients may also expose the recall, save_insight, and what_do_you_know prompts, and can open an owned memory through memory://item/{memory_id}. The server does not inspect chat history or local files; after an empty search or feed, it may issue one or two authenticated, read-only room or statistics probes to explain what the result did not cover.

1. A developer moving among Claude Code, Cursor, VS Code, and ChatGPT who wants technical preferences and project decisions to survive new sessions.
2. A team maintaining one codebase that wants assistants to share architecture decisions, ownership details, and previously learned engineering lessons through a room.
3. A user who repeatedly opens fresh chats and wants preferences such as frameworks, interface settings, or working habits recalled without restating them.
4. An engineering group that records deployment practices and past mistakes, then marks recalled advice as helpful or misleading to influence future ranking.
5. An MCP user who prefers a managed memory backend and does not want to operate a local vector database or memory service.

How do you install or deploy this agent?

Create a free API key at console.mnemoverse.com; the documented signup requires no credit card. On macOS, Linux, or Git Bash, verify it through a masked prompt so it is not placed in the command arguments:

printf 'Mnemoverse API key: '; read -rs KEY; echo
printf 'X-Api-Key: %s\n' "$KEY" | curl -s -H @- https://core.mnemoverse.com/api/v1/memory/stats; unset KEY

A JSON response containing "total_atoms" indicates that the key works. Install it for Claude Code at user scope, replacing the placeholder with the real key:

claude mcp add mnemoverse -s user \
  -e MNEMOVERSE_API_KEY=mk_live_YOUR_KEY \
  -e MNEMOVERSE_API_URL=https://core.mnemoverse.com/api/v1 \
  -- npx -y @mnemoverse/mcp-memory-server@latest

For Cursor, merge this into the global ~/.cursor/mcp.json; do not commit a real key in a project-level configuration:

{
  "mcpServers": {
    "mnemoverse": {
      "command": "npx",
      "args": [
        "-y",
        "@mnemoverse/mcp-memory-server@latest"
      ],
      "env": {
        "MNEMOVERSE_API_KEY": "mk_live_YOUR_KEY",
        "MNEMOVERSE_API_URL": "https://core.mnemoverse.com/api/v1"
      }
    }
  }
}

Restart the AI client after changing its configuration because MCP servers are loaded at startup.

How do you use this agent?

After installation and a client restart, ask the assistant to make an explicit write call:

Remember that my favourite TypeScript framework is Hono, and please call memory_write to save it.

Once storage is confirmed, open a new chat or session and ask:

What's my favourite TypeScript framework?

The assistant should call memory_read and answer “Hono.” If it does not, verify that the client was fully restarted and that mk_live_YOUR_KEY was replaced with a real key. In ongoing use, call memory_feedback to rate recalled memories, or create a room and issue scoped invitations when several users need a shared memory domain.

What are this agent's strengths and limitations?

How does this agent compare with similar options?

Key facts side by side with the most closely related agents.

How does FollowAgents rate this agent?

[![FARS review](https://followagents.com/badge/mcp-memory-server.svg)](https://followagents.com/en/agents/mcp-memory-server)

<a href="https://followagents.com/en/agents/mcp-memory-server"><img src="https://followagents.com/badge/mcp-memory-server.svg" alt="FARS review"></a>

Recent releases

Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.

1. v0.17.1 2026-11-05
   Release notes excerpt

   Documentation-only patch. Tool behavior, descriptions, schemas, annotations, permissions and dependencies are unchanged from 0.17.0. Documentation - Make the published Mnemoverse Memory plugin the first ChatGPT setup route in the README and agent installation guides, with custom MCP setup retained as an advanced alternative (#238). - Verify read access with an actual memory_stats result. For a user-requested save, inspect the write result and read the fact back from the same domain; visible tools or a conversational answer alone do not prove persistence. …

   View on GitHub ↗
2. v0.17.0 2026-10-30
   Release notes excerpt

   A MINOR: memory_read's tools/list entry gains a _meta field. The other changes are tool description text and documentation. Added - memory_read is kept loaded up front in Claude Code. Its tools/list entry carries _meta: {"anthropic/alwaysLoad": true}. Claude Code puts MCP tools behind tool search by default, so at the start of a task the read could be out of context entirely; this flag (Claude Code 2.1.285, https://code.claude.com/docs/en/mcp.md) opts one tool out. Only memory_read carries it; every other tool stays behind search. The flag takes effect once …

   View on GitHub ↗
3. v0.16.0 2026-10-23
   Release notes excerpt

   A MINOR: it adds an output field to two tools. Added - **memory_read and memory_list_recent name the shared rooms they did not search, as data and on a hit too.** Both declare a new optional output field, rooms_not_searched: {state, rooms?, note}, present only on a call without domain when the caller has rooms or the room list could not be checked. state is live (the rooms listed can be read by passing an address as domain), archived-only or unknown; rooms carries the live rooms in memory_list_rooms's row shape (the two now share one schema and one row …

   View on GitHub ↗

FAQ