Google AI Edge Foresight Brings Offline Meeting Notes to Macs

What Google shipped

Google has quietly released an AI meeting note-taker that runs entirely on your own computer. Google AI Edge Foresight is a free, experimental macOS app. It listens to meetings, transcribes them, turns a user's rough bullet points into finished notes and answers questions about what was said, all without a network connection.1211 Google released it on October 6, 2026, together with EmbeddingGemma 2, a new open embedding model, and it drew wider attention two days later when TechCrunch and The Verge covered it.1315 One outlet gives October 8 as the release date, which seems to reflect when the press coverage appeared rather than when the app went live.2

The app is built for Macs with Apple Silicon. Intel machines are not supported, and it is downloaded directly from Google rather than through the Mac App Store.913 Google's pitch is simple: your files, meeting audio and notes stay on your computer.12

Foresight is not likely to change the meeting-assistant market by itself. What it does show is that the cloud-first design behind most AI note-takers is now a choice, not something the technology requires. The pressure that creates for paid competitors matters more than the app.

How it works in a meeting

The workflow will be familiar to anyone who has used Granola. One side of the window holds the user's quick shorthand notes, and the other shows a live transcript built from the Mac's microphone and system audio.1 After the meeting, Foresight expands those fragments into fuller notes using what was actually said.12 TechCrunch's description of the split screen, with your notes beside AI-generated ones, explains why nearly every outlet calls the app a Granola rival.116

Foresight records at the system level and does not add a bot to the call. That means it works with Google Meet, Zoom or a face-to-face conversation without needing a separate plug-in for each platform.133 The upside is convenience. The downside, as one analysis points out, is that other people in the meeting never see a note-taker in the attendee list, so getting their consent is left entirely to the user.3 Another guide also tells users to get participants' consent before recording.

Foresight goes further than transcription. Users can build a knowledge base from PDFs, Google Docs, Microsoft Office files, plain text, Markdown and web bookmarks.6 A feature Google calls Live Assistance spots questions raised during a meeting and suggests answers drawn from that material as the conversation continues.76 Because the underlying model handles several media types, one plain-language search can cover images, documents, transcripts and notes together.10

The model stack, and a few discrepancies

The coverage mostly agrees on the technical setup but differs in the details. EmbeddingGemma 2 is a 740-million-parameter model that places text, images, audio and video in a shared search space. It handles retrieval, finding the right passage, file or moment in a recording.41 A Gemma 4 model does the generative work: writing the expanded notes and answering in chat.9 Heise describes the pairing as local retrieval-augmented generation, meaning the app searches your own data first and then writes an answer from what it found.9

Some reports simplify this. The Verge said Foresight runs on EmbeddingGemma 2 without mentioning Gemma 4, and Firstpost referred to an "on-device Gemma 2 model," which appears to be a garbled version of EmbeddingGemma 2.124 Which model handles transcription is also unclear. One account, citing an early demo, says it uses local Gemma 4 variants (E4B or 12B).7 The Gemma 4 model's memory requirements have not been published, though Google's figures put EmbeddingGemma 2 at about 191MB for text-only use and 567MB for the full multimodal version.10

A more practical discrepancy concerns recorded audio. The Verge said Foresight can transcribe meetings and audio files offline.12 A hands-on review found that the current version would not import an audio file for transcription; it could only record meetings live. Anyone hoping to process old interviews or voice memos should test this before relying on it.

On hardware, Google promises low battery drain thanks to compressed (quantized) Gemma models and memory-efficient search, but it has not published minimum memory or storage requirements.91

Why on-device matters for AI note-taking

The significance lies less in the features, which are now standard, than in where the processing happens. Most competitors, including Granola, Wispr Flow and Notion AI, send audio or text to cloud servers for processing.2 Google's own Gemini note-taking in Meet works the same way, while Foresight keeps the audio, transcript, notes and files on the Mac.13

This is most important in regulated or confidential work. Several outlets single out healthcare, legal and financial services as industries where uploading recordings creates compliance problems.23 AI Weekly's editors made the sharpest version of the argument: if local models can match cloud note-takers on quality, regulated industries lose their main reason for refusing these tools.6 A small-business guide made the same point in everyday terms, suggesting a dental office or accounting firm could use local models for notes and document search without client data reaching a vendor.7

The "it's free" angle matters too. Google says Foresight has no cloud subscription costs.10 Granola, by comparison, offers a free basic tier but charges $14 per user per month for Business and $35 for Enterprise.13 An experimental app won't replace those plans for teams that need sharing and admin controls. Still, a capable free local option from the company that builds the models makes it harder for anyone to charge for basic transcription and summaries.

The privacy claim has limits

The offline promise deserves a closer look than some coverage gives it. All of it is Google's own claim. One analysis notes that the launch post includes no independent audit and suggests confidential users watch the app's network traffic themselves during a test meeting.3 Others point to a further gap: if you connect Google Drive, those files have to be downloaded from Google's servers before they can be indexed locally. "Works offline" therefore describes where processing happens, not where every source comes from.13 The product page also links only to Google's general privacy policy and terms of service, not a privacy notice specific to Foresight.13

Output quality is the other open question. Expanding shorthand from a transcript can attach a note to the wrong part of the conversation, so names, figures and commitments need checking just as they would with a cloud tool.3 Since Google labels the app experimental, one outlet advises against making it the only record of an important meeting for now.10

A showcase or a real product?

The biggest question is whether Google plans to keep Foresight going. TechCrunch notes that the same team released an AI dictation tool built on local models in April, and that it never caught on widely. TechCrunch's reading is that Foresight may mainly be a demonstration of the Gemma models.11 Heise places it in a growing "AI Edge" series of Mac apps that already includes Google AI Edge Eloquent, a local transcription tool.9 Developer-focused coverage calls Foresight a reference design: a small search model paired with a larger writing model, running on a laptop with no server.103

My reading is that the showcase interpretation is right, but it doesn't make the app less important. Foresight shows what Google's open models can do inside a task people perform every day, and the same models are open for any developer to build on.8 Meanwhile, Google may be preparing a cloud-based note-taker in its Gemini apps that would work across video-calling services.4 One analysis sees a two-track strategy: on-device tools for privacy-sensitive work and Gemini agents in the cloud for complex enterprise tasks.2

In a category TechCrunch calls "competitive and crowded," now including note-takers from Wispr and Calendly and Superhuman's acquisition of Fathom, Google doesn't need Foresight to win.6 It only needs the app to change what buyers expect. Once a free Mac app can transcribe, summarize and answer questions with nothing leaving the device, cloud note-takers will need to justify sending recordings to their servers with clearly better note quality, collaboration and admin controls, not just habit. The open question is whether Gemma-class local models can match frontier cloud models on note quality, and whether Google will keep maintaining an app it calls experimental.3