A meter before an upload
A robot with a microphone has to earn some trust. So the first voice work kept a local microphone meter separate from cloud recording. Local mode shows audio levels without keeping or uploading clips. Cloud note capture needs pairing and a capture mode that has been switched on.
Keep each recording bounded
A voice note is capped at 30 seconds, with one clip buffer held in PSRAM. A signed upload hands the clip to the companion for transcription and note processing. An accepted upload is only the start: the companion still has to turn it into a note.
Recover when memory is tight
The September 19 work fixed a wake detector that could stay unavailable after a temporary allocation failure. The new development build prepares voice memory earlier and retries eligible detector failures at a bounded rate. One rule didn't bend: a failure never authorizes a fallback that uploads audio without wake detection.
Movement and listening have to cooperate
Motors are noisy, so capture now holds during head movement and for 750 milliseconds afterward, and a gesture can't start over an active note or upload. Host tests cover those coordination rules. Wake accuracy in a real household and motor-noise rejection still need physical checks, and direct speaker replies were switched off at this milestone.