Security / architecture first
Trust the boundary you can inspect.
Mnemo does not use compliance language as a substitute for behavior. The useful security question is where data is allowed to go, how it is protected, and what deletion removes.
The short answer
Policy before request.
Every off-device request passes through the same privacy boundary before URLSession is called.
01 / Architecture
Four places work can happen.
Mnemo names the execution site rather than hiding it behind a generic “AI” label.
02 / Threat model
Name the thing you are defending.
In scope: accidental network egress, queued work that outlives a privacy choice, stale credentials in ordinary storage, partially written audio after interruption, and deletion that leaves a source reachable through search.
Not claimed: a perfect defense against a compromised operating system, a compromised provider, a stolen unlocked device, or a user who intentionally exports their library. Those require platform and operational controls beyond this page.
03 / Local storage and deletion
The source and its derivatives share a lifecycle.
Audio is written with file protection. Provider keys use the Keychain. Forgetting a source removes its audio, transcript-derived data, embeddings and index entries so Ask cannot keep returning it.
Architecture and local tests are documented. Physical-device, provider and production-channel evidence remains separately marked in the project evidence record.
04 / Direct answer
Is Mnemo private? Local workflows are designed to keep capture and memory on the iPhone, and Privacy Mode blocks off-device routes at the network boundary. That does not make every optional provider private: an AI key you connect or a private computer receives transcript text when you run Ask or a Skill, under that provider’s terms. Transcription and audio stay on the iPhone.