Mnemoverse: managed memory over MCP, same memory in Cursor, Claude Code and VS Code

Hi all, Edward here, I build Mnemoverse.

It is a hosted memory API your agent reaches over MCP. You write a fact once (a preference, a decision, a lesson learned) and recall it in later sessions and from other tools: the same memory works in Cursor, Claude Code and VS Code with one key, and in ChatGPT via a Custom GPT action.

How it differs from the local-first memory servers shared here (Recall, Mnemosyne, PMB are all good projects): those keep everything on your machine, which wins on privacy and latency. Mnemoverse is the managed, cross-tool option, and it does not just store vectors: each write gets an importance score, related concepts link up, and when you report that a recalled memory actually helped, that feedback re-ranks future recall. So it learns which memories are worth surfacing.

Setup in Cursor: Settings → MCP, then

json {"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"}}}} ​

Free key at console.mnemoverse.com (1,000 queries/day, no card). MCP server and Python SDK are MIT; the engine is hosted. There is a research paper behind the memory model if you like that sort of thing: arXiv 2603.08965.

Honest caveat: if you want fully local and private, pick one of the local-first servers above. If you want one memory across your tools without running anything, that is what this is for. Happy to answer anything.

1 Like

Small update for anyone using this setup across Cursor and Claude Code.

If you added the memory server through Claude Code’s CLI and it only shows up in one project, the cause is scope: without an explicit flag the server is registered for the current folder only. Adding it with -s user registers it once for every project:

claude mcp add mnemoverse -s user -e MNEMOVERSE_API_KEY=your_key -- npx -y @mnemoverse/mcp-memory-server@latest

Cursor itself is not affected by this flag: there the same thing is controlled by where the config file lives. A global ~/.cursor/mcp.json makes the server available everywhere, a project’s .cursor/mcp.json keeps it local to that repo. If the whole point of your setup is one memory that follows you between tools, the global variants are the ones you want in both.

This came out of reviewing our own install path after the latest release; the server now also prints a hint to stderr on startup when the API key is wrong, so the silent-empty case is easier to spot.

1 Like