MCP (Model Context Protocol)
RememberOS speaks MCP over HTTP — point Claude Desktop, Claude Code, or any MCP client at it and your agent gets memory tools.
Endpoint#
POST https://rememberos.ai/v1/mcp (JSON-RPC 2.0)
Authorization: Bearer mv_…Tools#
| Tool | Args | Does |
|---|---|---|
search_memory | query, collection?, limit? | hybrid search |
store_memory | text, collection? | store a memory |
get_memory | memory_id, collection | fetch one memory |
list_collections | — | your collections + counts |
related_memory | id, limit? | nearest neighbours of a memory — discover connected observations |
what_changed | collection, limit? | where a newer fact superseded an older one ({current, superseded} pairs) |
collection_health | collection, stale_days? | total / stale / expired / supersessions + a needs_attention roll-up |
summarize_collection | collection, limit? | board-report briefing — {summary, themes, based_on} |
find_duplicates | collection, threshold? | near-duplicate clusters (read-only preview; deletes nothing) |
archive_stale | collection, older_than_days?, dry_run? | soft-archive stale memories (reversible; dry_run defaults true) |
The first four tools carry knowledge in and out; related_memory,
what_changed, and collection_health are discovery primitives — they let
one agent surface patterns, trends, and drift in knowledge a different agent recorded; while
find_duplicates and archive_stale let agents maintain the shared memory. That is
the point of a shared memory layer: the agent that writes an observation is rarely the one that acts on it.
Agent patterns#
- Recall-first: call
search_memorywith the user's intent before answering; cite what you find. - Learn-after: when the user states a durable fact or preference, call
store_memory. - Use one collection per project, or container tags per end-user (set via the REST API/SDKs — MCP tools keep the minimal surface).
For the one-paste setup prompt, see Install with AI. To wire agents from several vendors into one shared collection, see Connect every agent.