Your data, your control
No vague promises — here is the honest, hop-by-hop account. Your persistent memory lives in a store youcontrol, separate from any model vendor’s own memory. At answer time, data passes through model APIs; below is exactly which, and under what terms.
The text you ingest, encrypted at rest (AES-GCM) and PII-redacted first.
You control it — inspect, delete, revoke keys. Never shared across accounts.
The text you store, sent once to turn it into a search vector.
Not used for training under the API terms. Briefly retained for abuse monitoring, then deleted.
Your question plus the retrieved passages, to write the answer.
Not used for training under the commercial API terms. Briefly retained for abuse monitoring, then deleted.
The draft answer and passages, for the independent faithfulness check.
Not used for training under the API terms. Briefly retained for abuse monitoring, then deleted.
Every hop above is a commercial API that does not train on your data, so the hosted demo already keeps your content out of model training. If you want nothing to leave your own machine at all, run the engine yourself: the server is self-hostable (Docker), and you can point it at a local modelfor embeddings and answering. In that mode the store, the vectors, and the model are all yours — the strongest “your data, your control” setup.
This is a personal engineering project running on free-tier infrastructure, provided as-is with no warranty. The model providers’ own terms govern the data they receive; the summaries here reflect those terms at the time of writing and may change — verify the current terms if data handling is critical for your use.