claude mcp add --transport http agent-cabinet http://localhost:7879/mcp
claude mcp list

Add a project file at .cursor/mcp.json:

{
  "mcpServers": {
    "agent-cabinet": {
      "url": "http://localhost:7879/mcp"
    }
  }
}

codex mcp add agent-cabinet --url http://localhost:7879/mcp

# remember("The DB is Postgres, not MySQL.")
---
id: acme-project/the_db_is_postgres_not_mysql
type: domain_knowledge
scope: { source: acme-project }
---
The DB is Postgres, not MySQL.

# remember(..., basis="stated", links=["db-migration-plan"])
---
id: acme-project/the_db_is_postgres_not_mysql
type: domain_knowledge
scope: { source: acme-project }
basis: stated
links: [db-migration-plan]
---
The DB is Postgres, not MySQL.

# file(component, overwrite=True) -- promoted later, same id, by
# something that actually verified it
---
id: acme-project/the_db_is_postgres_not_mysql
type: domain_knowledge
scope: { source: warehouse-connection-check }
evidence: { confirmed_by: "SELECT version() -> PostgreSQL 16.2" }
provenance: { derivation: check_db_dialect, derivation_version: a1b2c3 }
---
The DB is Postgres 16.2, confirmed by direct connection.

recall() is a rebuild-every-call BM25 pass over the cabinet — the right default for the dozens-to-low-hundreds of entries a typical cabinet holds. No index file, no server, nothing that can drift out of sync with the Markdown files, which stay the sole source of truth.

A persisted, incremental index — dense/semantic retrieval alongside the lexical half — is a real upgrade once a cabinet grows into the thousands of entries. It isn't built yet; when it lands, it'll be a disposable cache derived from these same files, never a second copy of them.

• No multi-user access control. One writer, one directory — a person, or a small team sharing a git repo. Grants, SSO, and per-entry sharing are an enterprise concern for a separate product to layer on top.
• No confidence scores. basis (stated / inferred / observed / assumed) asks for something an agent actually knows — where a belief came from — not a calibrated probability it would have to guess at.