The trust policy is the part worth reading: memory is treated as advisory context rather than instruction authority, with a stated boundary between what gets read automatically and what requires an explicit write. That is the right position for a store an agent both reads and writes, and it is rarely written down. The connectors change the character of the thing — once repositories and Notion pages sync into project collections, recall stops being notes-you-took and starts being your actual work, which is also the point at which what you are handing a hosted service is worth thinking about.
A hosted context layer reached over Streamable HTTP. Decisions, preferences and project context are saved once and recalled from any MCP client, and optional connectors sync GitHub and Notion into the same store so repositories, commits, pull requests, issues and workspace pages are searchable alongside what you wrote by hand.
- The core loop: `remember` to save context with an optional `project:<slug>` collection, `recall` for semantic search across it, and `get_context` which returns a formatted block ready to drop into an agent prompt
- Browsing and retrieval — `list_memories` by collection, tags, type or visibility, `get_memory` for full content by id, `list_collections` for the scopes that exist
- Maintenance: `update` patches or appends to an existing memory, `forget` deletes one permanently, `memory_stats` reports by type and collection
- Four connector tools enabled per account for wiring sources from chat — `connect_source` starts the GitHub App install or Notion OAuth, `check_connect_status` polls it, `list_syncable_items` shows what is available, `set_sync_selection` chooses what to sync and triggers the first run
- Synced content landing in one collection per project, tagged and searchable next to manual memories
An account, authorized over OAuth 2.1 with PKCE — dynamic client registration is supported, so clients register themselves on first connect and there is no key to paste. Point the client at `https://mcp.memxus.com/mcp` with a streamable-http transport; nothing installs locally. The registry name is `com.memxus/memxus`, 1.3.2 in server.json. Self-hosting needs Node 20+, a Supabase project with the migration applied, and a set of environment variables including `MCP_PUBLIC_URL`, `SUPABASE_URL`, `SUPABASE_SERVICE_ROLE_KEY`, `OAUTH_CLIENT_ID` and `ALLOWED_REDIRECT_URIS`, with `OPENAI_API_KEY` optional for vector search embeddings. Licensed AGPL-3.0, which means running it as a network service obliges you to publish your source.
One command — npx -y smithery mcp add memxus/memxus --client cursor --force