Skip to content

About

Claude Code plugin: persistent Mnemoverse memory over MCP with OAuth sign-in, plus the CC0 agent-memory-discipline skill.

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

Mnemoverse Memory

HOL trust score

Persistent memory for AI agents, shared across tools. One account gives Claude Code the same long-term memory it has in Cursor, VS Code, and any other MCP client: write a memory in one tool, recall it in another.

This plugin connects the hosted Mnemoverse memory server over MCP and adds the discipline that makes memory actually get used: recall before acting, save durable decisions and corrections afterwards.

Install

One line, no directory and no review queue:

/plugin marketplace add mnemoverse/claude-plugin
/plugin install mnemoverse@mnemoverse

The first command registers this repository as a plugin marketplace; the second installs the plugin from it. Updates arrive with /plugin marketplace update mnemoverse.

To try it from a local checkout of this repository instead:

claude --plugin-dir ./

Other hosts: OpenClaw and Hermes Agent

Both hosts install this repository as a plugin and load its skills, but neither starts a sign-in for a server that comes from a plugin. Add the hosted server once as the host's own entry, under the same name, and sign in there.

OpenClaw

openclaw plugins install mnemoverse --marketplace https://github.com/mnemoverse/claude-plugin
openclaw plugins inspect mnemoverse
openclaw mcp add mnemoverse --transport streamable-http --url https://mcp.mnemoverse.com/mcp --auth oauth
openclaw mcp login mnemoverse

mcp login prints the sign-in URL instead of opening a browser, so open it yourself and approve the scopes; it waits on a loopback callback, which the Mnemoverse sign-in accepts on any port. openclaw mcp probe mnemoverse lists the tools once you are signed in. Full walkthrough: Install by host, OpenClaw.

Hermes Agent

hermes plugins install mnemoverse/claude-plugin --no-enable
hermes plugins list
hermes plugins enable <name printed by hermes plugins list>

Then add the server to ~/.hermes/config.yaml:

mcp_servers:
  mnemoverse:
    url: "https://mcp.mnemoverse.com/mcp"
    auth: oauth

and run hermes mcp login mnemoverse (Hermes opens the browser and waits on a loopback callback). Check with hermes mcp test mnemoverse. Keep the name mnemoverse: the native entry shadows the plugin's copy, which cannot sign in. Full walkthrough: Install by host, Hermes Agent.

Google Antigravity

The antigravity/ folder is a self-contained Antigravity plugin with the hosted server and the bundled skills:

agy plugin install https://github.com/mnemoverse/claude-plugin/tree/main/antigravity

The plugin installs enabled, and Antigravity opens the browser for the sign-in when the server first connects; there the server is named mnemoverse_mnemoverse. Details in antigravity/README.md. Installing the whole repository with agy plugin install https://github.com/mnemoverse/claude-plugin also works: Antigravity reads the server from the root mcp_config.json and converts the Claude Code commands into skills.

First run: sign in once, no API key

The server at https://mcp.mnemoverse.com/mcp uses OAuth 2.1 with PKCE. In Claude Code, installing the plugin does not open anything by itself: run /mcp, select the Mnemoverse server (from this plugin it is listed as plugin:mnemoverse:mnemoverse) and choose Authenticate, then sign in to your Mnemoverse account in the browser (free tier at console.mnemoverse.com, no credit card) and grant the requested scopes, which include memory:read and memory:write. There is no API key to paste. To disconnect, remove the server from your client; resetting your password in the console can also disconnect all connected apps at once. The bundled setup skill walks the agent through the current host's sign-in flow if anything goes wrong.

After sign-in, ask the agent to check the connection with memory_stats. A successful stats result confirms read access, including when the account is empty. Tools appearing in a list are only discovery; an error or empty response does not confirm the connection. This check does not write a memory or verify writes. If you also asked the agent to save or recall something, that task needs its own successful tool result.

What you get

Three commands:

Command What it does
/mnemoverse:remember save a durable fact, decision, or correction, one fact per memory
/mnemoverse:recall search memory before acting, with provenance, surfacing contradictions instead of hiding them
/mnemoverse:memory-status counts, domains, and the newest memories

Seven skills:

  • agent-memory-discipline (CC0, backend-neutral): standing rules for when to recall and what is worth saving. It names no vendor and works on any memory backend; it ships here so the connected memory is actually used.
  • setup: guides the current host through connecting and troubleshooting the MCP server.
  • memory-search: searches memory before answering, then rates which recalled memories helped.
  • memory-save: saves a decision, preference, correction or lesson as one self-contained fact, never a secret.
  • memory-catch-up: lists what was added since a given time, in personal memory or in a shared room (beta).
  • memory-onboard: seeds memory with a project's conventions and decisions from its README, CLAUDE.md and AGENTS.md.
  • mnemoverse-docs: for developers adding Mnemoverse memory to their own app or agent; picks the form and points to the docs page.

The tools come from the remote server itself: reading and writing memories, a newest-first listing, stats, usefulness feedback, shared rooms (beta), and a vault alias listing (aliases only, values are never returned). The current list, with descriptions and schemas, is at the server's card.

Try these

  1. Remember that this project uses pnpm, not npm. We decided this after a lockfile conflict on 2026-08-01.
  2. Before you touch the billing module, recall what we have decided about it in past sessions.
  3. /mnemoverse:memory-status after a week of use, to see what your sessions have been saving.

What gets stored, and what does not

The plugin stores only what you or the connected agent explicitly save through the memory tools: single facts, decisions, corrections. It does not record conversations, does not read your chat history, and the bundled skill refuses to persist passwords, API keys, tokens, payment data, MFA codes, government IDs or health records.

Memory content lives in your Mnemoverse account. Privacy policy: mnemoverse.com/docs/legal/privacy-policy.

Troubleshooting

  • No sign-in prompt in Claude Code: run /mcp, select plugin:mnemoverse:mnemoverse and choose Authenticate. If the server is absent from Claude Code's configuration, add it for all your projects with claude mcp add --transport http --scope user mnemoverse https://mcp.mnemoverse.com/mcp; it then appears in /mcp as mnemoverse, and you authenticate it the same way. Other hosts use their own sign-in flows described above.
  • 401 after it used to work: authentication failed. Use the same host's sign-in flow for the same server; in Claude Code, run /mcp and choose Authenticate or Re-authenticate as offered. Then check memory_stats again. A 403 means access was denied, and a timeout leaves the connection unverified; neither proves an expired token.
  • malformed_token in logs: the bearer token is neither an OAuth token from the sign-in nor a Mnemoverse API key, for example a key copied without its mk_live_ prefix, or a placeholder left in a config. Since 2026-10-19 the remote endpoint accepts either: a host that cannot complete the sign-in can send an API key as Authorization: Bearer mk_live_... (details). A revoked or unknown key gets invalid_api_key instead.

More: docs.

Why .mcp.json says http and mcp.json says streamable-http

Both files point at the same server, but each is read by a different host with its own vocabulary for the transport. .mcp.json is Claude Code's own project MCP config, where type: "http" is the canonical value (code.claude.com/docs/en/mcp.md); streamable-http is accepted there only as an alias. Root mcp.json follows the Agent Plugins 1.0.0 schema instead, whose type enum is stdio | streamable-http | sse and rejects "http" outright (agent-plugins.org/schemas/1.0.0/mcp.schema.json). kimi.plugin.json, the Cursor and Gemini manifests and the Antigravity mcp_config.json files add more dialects again (see scripts/fixtures/). scripts/check-manifest-drift.mjs guards every declaration of the server URL against drift while keeping each file's dialect intact; see its CI job, manifest-drift-guard.yml, for what it checks.

Support

Issues in this repository, or support@mnemoverse.com.

The open-source MCP server package (MIT) lives at mnemoverse/mcp-memory-server. A star there helps other builders find the project.

License

Plugin: MIT. The bundled agent-memory-discipline skill is CC0-1.0 and lives at mnemoverse/agent-memory-discipline.

About

Claude Code plugin: persistent Mnemoverse memory over MCP with OAuth sign-in, plus the CC0 agent-memory-discipline skill.

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages