Permanent, provenance-aware memory for your agents.
SyntheticMem is a pure MCP server (with a plain-REST mirror) that gives any AI agent shared, semantic recall that outlives the session. Every memory carries its source, its confidence, and the condition that would make it wrong — so your agents verify instead of blindly trusting.
Explore the docs
Getting started
Provision an agt_ key, mount the server, and run your first remember + recall over both MCP and REST.
Concepts
The memory model: provenance, falsifiability, the fact → hypothesis → retired machine, ephemeral vs permanent, and “trust, but verify”.
Tools reference
The six MCP tools (plus ping), each with its input schema, an example exchange, and the equivalent REST endpoint.
Connect your agent
Claude Code / Desktop, opencode, grok, and raw HTTP — the exact connect strings for each client.
REST API
A plain-REST mirror of the tools under /api/v1, with interactive Swagger UI at /api/docs.
How it works
MCP over Streamable-HTTP, pgvector (vector(768) + HNSW), local ollama embeddings, and the prune scheduler.
The one-liner
Mount the server on any Streamable-HTTP MCP client with a per-agent bearer key:
claude mcp add --transport http syntheticmem https://syntheticmem.com/api/mcp \ --header "Authorization: Bearer agt_…"
The agent forgets. The memory doesn’t.