Google Tests a Mac Notes App That Transcribes Meetings Offline
Google has released Google AI Edge Foresight, an experimental and free note-taking app for macOS that transcribes meetings and audio files entirely on-device. Current builds are optimized only for Macs with Apple Silicon, Apple's own processors. The Verge
TechCrunch described the app as a local-first competitor to Granola, meaning it keeps data on the computer instead of sending it to the cloud. TechCrunch The Verge, in Oct. 8 coverage, noted TechCrunch's earlier report of the release. There is no cloud transcription. There is no upload queue. Capture and retrieval stay on the machine.
Google's developer documentation describes Foresight as "completely offline-first," with both the Gemma model, Google's family of open models, and the local vector embeddings engine, the software that turns content into searchable numbers, running on user hardware without internet. Transcription, embedding, indexing and query all execute locally. For meeting audio, that removes network dependency and keeps raw audio and transcripts out of a server-side pipeline.
The transcription and retrieval path is built around EmbeddingGemma 2. Google describes EmbeddingGemma 2 as an open, lightweight multimodal model, meaning its weights are public and it can handle more than one type of input. Google Blog The developer-focused writeup lists it as a 740M multimodal model, open-weight, and designed for private, ultra-low-latency on-device retrieval, or fast private search on the device. Google Developers Blog
EmbeddingGemma 2 enables semantic search, routing, and retrieval fully on-device, and brings robust capabilities directly to edge hardware, according to Google. In practice for Foresight, that means spoken content is converted to searchable embeddings, numerical summaries of meaning, in a local vector store on the laptop. Users query past meetings by meaning, not filename or keyword. Routing determines what should be handled locally and what passage should be surfaced.
The interaction model is light. For Mac, the app turns scribbles into full notes using offline recordings. 9to5Google Short input during a call becomes the retrieval key, the short prompt that finds the right material. The full transcript and generated notes do the rest.
The broader context here is that Foresight functions as a reference implementation for the Google AI Edge stack, not only as a utility. Read it that way. It shows an embeddings model, a local vector engine, and an application layer integrated tightly enough to handle continuous speech, speaker turns, and ad hoc queries within laptop power and thermal limits. Latency, the wait for a result, matters here. So does persistence, keeping the index usable over time. A note-taking app exercises both relentlessly.
In my view, the choice of meetings as the workload is pragmatic. Enterprise and technical teams already live in recorded calls, yet many avoid cloud transcription for legal, client, or security reasons. Local-first capture lowers that friction. It also shifts failure modes. Instead of worrying about retention policies on a SaaS server, users manage disk, index size, and device migration. Those are familiar tradeoffs for engineers. They are easier to reason about than opaque cloud handling.
Worth flagging is what this suggests for on-device retrieval design. A 740M open-weight embeddings model is small enough to ship, yet capable enough to support semantic search across a growing personal corpus, a collection of your own material. If that holds up under daily use, the pattern extends beyond notes. Support logs, field recordings, lab audio, and research interviews all fit the same pipeline. No new infrastructure required.
What makes this small experiment worth watching for practitioners is that it will succeed or fail on transcription accuracy, retrieval quality, and resource footprint. Those metrics are measurable. And if local-first meeting notes become routine, private AI stops being a principle and becomes a workflow.
