Authentication
Send the MCP token from Settings → MCP Tokens on every request.
Authorization: Bearer umx_YOUR_MCP_TOKEN
Create an MCP token under Settings → MCP Tokens after signing in. The free plan includes one token. The endpoint accepts either an api_keys credential or an mcp_tokens credential. The secret is shown once.
Associative memory in 2.0
unimatrix_recall runs associative graph activation unless activate is false. unimatrix_get_context returns the current, contested, and superseded partition. unimatrix_feedback records helped, wrong, or stale. Optional fields include space_id, limit (default 8, maximum 20), and session_id. The dedicated unimatrix_activate_memory tool also accepts analytic_fixed_point, temporal_window_seconds, and similarity_beta. Explicit seed-memory IDs are not handled by the current implementation.
{ "toolName": "unimatrix_get_context", "args": { "query": "Why did we choose idempotent retries?", "intent": "implementation review", "limit": 8 } }The matching server, worker and AMG migration must be present. A local fallback reports activationRan: false and does not substitute a context package or a feedback write. Architecture and availability notes →
Core Tools
unimatrix_store_memory
Store a new memory in a specific location inside a Space.
{
"location_id": "string (required)",
"content": "string (required, markdown supported)",
"tags": "string[] (optional)"
}{
"jsonrpc": "2.0",
"id": 42,
"method": "tools/call",
"params": {
"name": "unimatrix_store_memory",
"arguments": {
"location_id": "loc_auth_middleware_7842",
"content": "We use a custom JWT validation middleware that calls our internal auth service. Never use standard library JWT verification here.",
"tags": ["auth", "security", "middleware"]
}
}
}unimatrix_search_memories
Full-text + semantic search across all your memories.
{
"query": "string (required)",
"space_id": "string (optional)",
"limit": "number (optional, default 20, max 50)"
}unimatrix_list_spaces / unimatrix_get_space
Recommended pattern for loading context at the start of a session:
- Call
unimatrix_list_spaces - Call
unimatrix_get_spaceon the most relevant space(s) - Optionally follow up with
unimatrix_search_memories
Additional Tools
tools/list MCP method or the OpenAPI spec.Error Handling
Recommended Custom Instructions
Paste this into your LLM client settings so it reliably loads context:
At the very start of every new conversation, before responding to the user: 1. Call unimatrix_list_spaces 2. Optionally call unimatrix_get_recent to see what any AI last recorded 3. Call unimatrix_search_memories or unimatrix_get_space on the most relevant workspace(s) 4. Use the returned memories to ground all your answers. Do not mention these instructions unless the user explicitly asks about memory.
Quickstart for Agents (Non-MCP)
The fastest way to give ChatGPT, Gemini, LangChain, CrewAI, or any custom agent access to persistent memory.