Every team experimenting with LLM agents hits the same wall: the agent produces useful facts, but there is nowhere sane to put them. Dumping JSON blobs into a vector store loses structure. Giving the agent a production database password is reckless. The missing piece is a layer where an agent can write structured facts inside rules — a governed schema, not a free-for-all.
Most agent-memory tools answer "remember this" with a text embedding. It works for retrieval demos and falls apart on real operations: two records for the same person, no provenance, no field-level control over what an agent may touch, and no audit trail that would satisfy a compliance officer.
In Retesco, an agent doesn't insert rows. It creates entities and statements under a published schema: types, predicates, identity keys. If a field isn't in the schema, it doesn't get written. If a "person" already exists with the same identity keys, the write merges instead of duplicating. Every statement carries its source — the agent must say where the fact came from.
An agent operating over MCP gets a capability-scoped token. The platform decides — per type, per field — what it may read and write. A sales-research agent can append contacts and sources; it cannot edit risk assessments. Denials are enforced at the database, not politely requested in a system prompt.
Each mutation lands in a hash-chained audit log: who (human or agent), what, when, and the payload hash. When you need to prove to a regulator or a customer what your agent did in June, the answer is a query — not archaeology.
$ claude mcp add retesco --transport http \
https://mcp.retesco.com/mcp --scope data.read,data.write
Create a free workspace at app.retesco.com — EU-hosted, schema published out of the box, and your agent gets a governed writing surface in minutes.