What the jacket hears,
and where it goes.

Two lavalier microphones sit in the collar. One faces forward, one faces back.

Your body is between them and acts as a baffle, so the front and back levels differ. Only the person facing you comes through at a usable level. Someone across the room is too quiet on both.

The back microphone is what makes that judgement. A voice that arrives louder from behind you is not someone you are in conversation with, and the app discards it instead of keeping it.

This is set by where the microphones sit, not by a software setting. It is not a room mic.

In California, where every Founder Edition is delivered, recording a private conversation needs the consent of everyone in it. The jacket is built on that assumption, not around it.

When the app hears a voice it has not heard before, it waits. You tell the person, out loud, the way you would anyway, and the app listens for their answer. Nothing is transcribed until it hears them agree.

Nothing has to interrupt you for that to work. The asking is part of the conversation, not a prompt on a screen.

If they do not agree, nothing from that voice is kept.

If you were recorded and want your data gone, ask the wearer. It is the same as asking someone to delete a photo they took of you. They delete it on their Mac and it is gone. There is no other copy.

Audio, voice embeddings, recordings, and transcripts never go to Isper or to any server.

The phone records. In the evening it hands the day to your Mac over a direct device-to-device link. No network, no account. You compare six digits on the two screens to confirm they are talking to each other.

The phone deletes its copy only after the Mac has proved, byte for byte, that it holds the same recording. There is never a third copy.

Transcription, speaker separation, and the writing run on the Mac.

What does touch the network: model downloads and app updates. That is the whole list.

The one way anything leaves is a choice you make: connecting an AI to the library. That is the next question.

The Mac app speaks MCP, so an AI you already use can answer questions from your library. Nothing is connected until you connect it.

When you ask, your question and the parts of the library needed to answer it go to that model, and nothing else does. The recordings stay on the Mac.

With a model running on your Mac, nothing leaves the Mac.

With a cloud model, what you sent is covered by that provider's terms, not ours. You chose the provider, and you can disconnect it at any time. The library is untouched either way.

This is deliberate. Once the day is processed, what you do with it is yours to decide.

On your phone: recordings, until they have been handed to the Mac. Then they are released.

On your Mac: the recordings, the transcripts linked to the person who said them, and voice embeddings. An embedding is a small numeric signature of a voice, so the app can recognise the same person weeks later. That is the mechanism.

There is no cloud copy, by design. Lose the phone before the handoff and that day is gone, the same as a photo you had not backed up.

The Mac writes a page per person from what it heard. Like any transcript, a name or a detail can be off.

Review and Corrections in the Mac app show what was written, from which conversation, and a fix sticks.

The recording it came from stays on the Mac, so anything written can be checked against what was actually said.

The jacket has no battery and no radio. Unplugged, it is physically incapable of recording.

You can also stop it listening in the app, which is quicker when the jacket is under a coat.

The cable is the one you can show someone. A software mute is a promise; an unplugged cable is a fact another person can see for themselves.

It is on your Mac and your phone, nowhere else. Delete the library in the Mac app and delete the iOS app, and it is gone.

Individual people and conversations can be deleted inside the Mac app.