Echo

Echo turns a day of conversations into something you can ask questions of, without handing your voice to the cloud. A watch keeps only speech and streams it to the phone, which transcribes and indexes it entirely on-device. Lookups ("When is the plumber coming?") are answered locally in about 30 milliseconds, and promises are collected into a checklist with plain rules. Questions that need synthesis ("What did I promise people today?") go to Meta's Muse Spark through the Model API: Echo plans the smallest set of text snippets, sends text only and never audio, and asks for a structured answer whose citations map back to the original audio on the phone. A privacy ledger records every question and the exact text and bytes sent. You can review each request before it leaves, mask people's names, turn the cloud off, export the ledger, or forget the day, and you can ask out loud, from the phone or the watch: the question is understood on the phone, answered, read back and kept in a history on the phone. In our demo day, 38% of the recording was never stored and Muse received 27 snippets (4.9 KB) for the promises question. The microphone is simulated with a pre-recorded day, and the on-device Muse tier is a design; everything else runs for real.