Plenty of memory servers embed and retrieve. Here a memory has a state that moves: confidence rises when something else corroborates it, falls when it does not, and unused memories fade. That is what stops a store filling with things that were true once. Note the split in memory types — a procedure or a script reference is static and should not decay, and choosing the wrong type is the mistake that will make recall behave oddly.
A managed memory service reachable over MCP or REST. Stored memories are embedded, mined for entities and relationships, and retrieved by four strategies at once — vector similarity, temporal recency, keyword search and graph traversal — with the system learning which to weight for each kind of query.
- `store` saves a memory, generating embeddings and extracting entities automatically
- `recall` searches with multi-strategy retrieval and cognitive scoring, ranked by an activation model drawn from cognitive science that factors in recency, access frequency, entity connections and validated confidence
- `update` modifies a memory and re-embeds it if the content changed; `forget` removes one by id or by closest match to a query
- `graph` traverses entity relationships by name and depth
- `status` reports memory counts, confidence distribution and detected knowledge gaps; `snapshot` gives a formatted overview organised by type
- Memory types change how a memory is managed: `general_knowledge` and `user_knowledge` evolve, gaining and losing confidence over a lifecycle, while `callable_scripts`, `work_project`, `cross_reference` and `learned_procedure` are static reference
- Every action is also an HTTP endpoint, so the same store is reachable from code that is not an MCP client
An account at adaptiverecall.com, which issues a server URL and an API key; the free tier is enough to start. Add it as an HTTP MCP server with the key in an Authorization bearer header — Claude Code, Codex, Cursor and Gemini CLI configurations are all documented. The registry entry names ADAPTIVE_RECALL_API_KEY as a required secret and the npm package adaptive-recall, version 1.0.1.
One command plus a key — npx -y mcp-remote https://YOUR_SERVER_URL/mcp, then supply credentials