What we can and cannot see, and why the architecture, not our word, is the answer.

The key model

Architectural constraint, not a policy position.

Your information is encrypted with a key derived from your password before it is written to disk. We do not have the key. This is an architectural constraint, not a policy position. The analogy that works is Signal. Signal does not have the keys to your messages either. That is not a statement about Signal's trustworthiness. It is a statement about what Signal's servers are structurally capable of. A government request, a hosting breach, or a Mandaire acquisition all produce the same result for the requester: encrypted blobs they cannot read.

Two questions come up consistently. The first is whether Mandaire reads your email to train a model. It does not. Your information is ingested to build your personal picture and used for nothing else. The LLM that reasons over your context is the one you bring: Claude Max, ChatGPT Plus, Gemini. Mandaire is the memory layer; the inference is yours. Neither sees more than the specific context needed for the specific question.

The second is what happens if Mandaire shuts down. You get your data back. Full export of your email index, decision ledger, taste memory, and relationship graph in portable formats. The encryption module is open-source under AGPL, which means any competent engineer can read your archive without Mandaire's involvement. No lock-in is an architectural commitment rather than a preference, because one of the bets the system is built around is that it should be useful without being load-bearing.