What it does
This local MCP server gives coding agents such as Claude Code, Codex and Cursor a shared project memory. As agents work, they record what they learn, and each claim keeps the host that asserted it, the principal it was made for, the workspace, its source, the observation time, a confidence value and a trust level computed by the receiving node. Contradictory claims remain side by side for resolution rather than the later one overwriting the earlier. The observation log is append-only, so a correction is a new observation, and entity merges and schema changes go through a proposal that shows what the verdict would overturn. A workspace leaves the machine only when it is listed, and peers are authorized per workspace.
It suits a small team that wants one project's memory on each machine, shared with teammates without a vendor cloud. It also fits setups where several agents, machines or people write into the same memory and disagreements need to be traceable. Credential-shaped text is refused at ingest, so secrets do not enter the store. With Claude Code, an optional hook can ask the agent for the reasoning behind a commit it made, and the hook never writes anything itself.
The server ships as one binary with an embedded store and needs no account. A connect command registers a bridge with each supported app, and a desktop app on macOS offers a settings page for the same tasks. Semantic retrieval uses a local multilingual embedding model of about 224 MB that is fetched once; until then retrieval is keyword only. Once a tool call has loaded the model, the daemon uses about 1 GB of memory. Removing the tool leaves the stored knowledge in place unless it is deleted separately.
Limitations and requirements
Semantic retrieval needs a local embedding model of about 224 MB, fetched once; until it is present, retrieval is keyword only. Once loaded, the daemon uses about 1 GB of memory. Claude Desktop accepts local servers over stdio only. Network access is used only when a peer is configured. Removing the tool does not delete the stored knowledge unless that is done separately.
Safety
Our team has not reviewed this resource. Only automated rules checked it; resources with high-risk signals are not published. These are signals, not a guarantee that installing it is safe. Try it in a sandbox project first.
- Runs code or changes an assistant's behaviour
Sources
- Official MCP Registry first seen 13 Nov 2026 · last seen 13 Nov 2026 · its page on the source
supragnosis
Memory for AI agents that remembers who said it.
v0.11.0As shown on registry.modelcontextprotocol.io when we last read it