Google AI Edge Foresight Brings Offline Meeting Notes to the Mac

A quiet launch with a loud positioning

Google AI Edge Foresight appeared on 8 October as an experimental, free macOS app for meeting notes and voice transcription, optimized for Apple Silicon, and the notable part is what it refuses to do: nothing leaves the machine. Recording, transcription, and the AI summaries all run on-device, and the app adds local document search on top.

The workflow is aimed at how people actually take notes in meetings: you jot quick shorthand while Foresight listens in the background, and afterwards the app pairs your bullet points with AI-generated summaries keyed to what was said. Coverage has described it as Google's local-first competitor to Granola, the paid AI notetaker, with the differentiators being the price (free) and the architecture (fully offline).

Why an on-device notetaker is harder than it sounds

Meeting transcription that never touches a server has to solve three things at once: real-time speech recognition on a laptop's compute budget, summaries that are actually grounded in what was said, and document search that works across formats without uploading anything. Foresight leans on Google's AI Edge stack, the same tooling behind the EmbeddingGemma models that run multimodal embeddings in a few hundred megabytes of RAM.

The privacy case writes itself for the meeting context specifically: recorded conversations are the most sensitive asset most professionals generate, and the notetaker category has repeatedly stumbled on consent and retention questions when audio flows to a cloud. An architecture where the recording never exists anywhere but your machine removes that class of incident.

The competitive signal

Google shipping a free, offline Granola competitor says two things about the market. First, on-device inference on Apple Silicon has crossed the usefulness threshold for sustained speech work, not just party tricks; the same week, the memory crunch that makes AI-capable machines more expensive to build became the industry's supply-side story. Second, when the largest cloud AI vendor starts competing on the absence of its own cloud, local-first stops being a niche selling point and becomes table stakes for anything touching private conversations.

For anyone choosing a notetaker now, the practical checklist from Foresight's design: where does the audio live, what leaves the device, is the summary grounded in the transcript, and can you verify all of that. A free first-party app meeting those bars is new; the questions are not.

It also sets a testable baseline for the category. An offline notetaker makes two claims a reviewer can check without trusting the vendor: disconnect the network mid-meeting and confirm transcription continues, and inspect the process's network connections over a full workday and confirm nothing recurs. Apps that route audio through a cloud API fail both tests by design, whatever their privacy policies promise, and Granola-class competitors that want to argue for their cloud features now have to explain what those features buy that local inference does not.