Consent, retention and the roster are three answers to one question: what this tool is allowed to keep, and for how long. Consent decides whether a recording starts at all, retention decides how long the audio outlives the transcript, and the roster decides what is stored about the people in it.
Where the data goes
Sessions are ordinary files in ~/Documents/Ambient unless you choose another
folder. Ambient does not encrypt them separately from your Mac’s storage.
Cloud folder sync and backups can copy them elsewhere. Exports and MCP clients
can also share transcript text beyond this Mac; see MCP access.
The live assistant, when you turn it on,
hands the transcript to the agent you run in Claude Code, which sends it to
that agent’s model, and keeps a log of what it said in the session’s
assistant.jsonl.
The recording prompt asks you, not the other participants. Obtain appropriate permission before recording. Ambient does not join the call or announce the recording to the people in it.
Recording states
The menu bar is the consent surface — the window can be closed, and answering about a call never needs it open — and it is a state machine:
stateDiagram-v2
direction LR
[*] --> Idle
Idle --> Armed: a watched app starts audio
Armed --> Recording: Record this call
Armed --> Idle: the app it armed for goes quiet
Armed --> Idle: Not this one, remembered until that app is quiet
Armed --> Recording: Start Recording, from the menu
Idle --> Recording: Start Recording, from the menu or the window
Idle --> Recording: a watched app starts audio<br/>and ask_before_recording is off
Recording --> Stopping: Stop
Recording --> Recording: Stop failed
Stopping --> Idle: audio written, transcript queued
Stopping --> Failed: the capture failed
Failed --> Idle: dismissed
Failed --> Armed: a watched app starts audio
note right of Armed
An empty apps list arms on nothing.
Quit from Recording or Stopping, or
with transcripts still queued, is
deferred, not refused: the app finishes
the outstanding work and then quits.
end note
Three of those edges are worth reading twice. A stop that fails leaves the
phase at Recording, because that is the honest state — the worker really is
still recording, so Stop stays enabled and Start stays disabled rather than the
app advancing past a capture it never stopped. Failed is watched exactly
like Idle: a failure nobody has dismissed does not stop the next call being
noticed. Transcription runs beside these recording states: completed turns
appear during capture, and the queue finishes remaining text and speaker labels
after Stop. A new recording can start while the previous session finishes.
The menu shows what the queue is doing and how
many sessions are waiting, and the icon is an hourglass until it is empty
(ADR-0015).
Consent: the armed state
With apps set, the menu bar watches for one of them producing audio and moves
to a fourth state, Armed — the menu then offers Record this call or Not
this one. Declining is remembered against that bundle id and no other, so
saying no to Teams says nothing about Zoom, and it is forgotten once that app
stops producing audio, so it does not opt out of the next call either. Standing
down from Armed is keyed the same way: it is the app Ambient armed for going
quiet that ends the wait, not the machine falling silent, so a music player
left running all day cannot hold the prompt open after the call has ended.
An empty apps list watches nothing. That list already means “capture all
system audio”, and arming on any sound at all would flap at every notification
chime; the Start item stays the way in.
The check runs on the existing refresh timer, every eighth tick — roughly every four seconds. Enumerating audio processes twice a second would be waste for something that changes when a human joins a call.
There is no system notification for the armed state. The menu bar is the
recording control: an unmistakable state you can
see without clicking. A real UNUserNotificationCenter prompt is additive and
should be scoped on its own rather than smuggled in here.
Retention
Audio is the most sensitive artefact here and the least useful once the
transcript exists, so it ages out on a schedule while the text does not.
audio_retention_days defaults to 7; 0 deletes the wavs as soon as the
transcript is written, forever keeps them. ambient delete is a separate,
deliberate action that removes the whole session directory — retention only
ever removes audio. The sweep runs at the end of every
recording — the app is running whenever a recording happens, so no launchd
agent is needed — and it never touches raw.jsonl, edits.jsonl,
session.json or transcript.md.
Live transcription temporarily keeps an additional native-rate PCM copy of
each track under audio/. Those working copies are removed after the finished
transcript is written. A small live-asr.json checkpoint remains so retries
preserve lines already read by the window or an MCP client.
Two guards matter. A session with no transcript.md is never swept: it is
still being written, and its audio is the only copy of what was said. That guard
has a consequence worth stating plainly — a recording that failed also has no
transcript.md, so its audio is kept indefinitely rather than ageing out. Those
directories are the ones to check by hand if you care about the retention
window holding. And
ambient diarize refuses on a session whose audio has been swept, because
a re-run reverts the previous labels before it discovers there is nothing to
read — which would silently unlabel every speaker nobody had named by hand. The
window says the same thing before you can ask: on a swept session Separate
voices is disabled and carries the reason under it, rather than offering a
button that would only refuse.
Who’s who
ambient roster add <name> keeps a list of the people you record with, in
roster.json beside the config. Under the transcript of whichever session is
selected in the window, each speaker diarization could not name gets a row
carrying the first thing that voice said, and a dropdown of roster names puts a
name on every line of that speaker. It is any session and not only the most
recent one: the naming strip follows the selection.
The roster stores names, not voiceprints. Ambient does not use it to recognize a person across sessions. You choose which name belongs to a speaker; check the result before sharing the transcript.
Naming from the window writes the same edit ambient name does, so it is
recorded as the user’s and survives a re-diarize.