Not "we promise not to look."
We are incapable of looking.

Two local ONNX models plus checksum-validated pattern matching run entirely on your machine (~70ms per message). Nothing is sent to the cloud to be "checked": the check is local.

The gateway that proxies AI traffic cannot log message content: it is a compile error, not a policy. It sees only user, model, token counts and timestamp.

Everything lives in an encrypted local database. Keys derive from a passphrase Amnesia AI® never sees; a Recovery Kit is issued at setup. No escrow, no back door: we cannot reset it, and that is the point.

On any error or ambiguity, Amnesia AI® blocks the send, tells you why, and records a receipt. It never falls through to sending unprotected data.

Every send produces a hash-chained receipt: time, model, what was masked, state, hash. One click verifies the whole chain. Proof that every AI use was masked, for you, your firm, or an auditor.

Encrypted export and import are free on every plan, with a standalone open verifier. Your data leaves with you, encrypted, whenever you want.