Google AI Edge Foresight Brings Offline Meeting Notes to Macs

What Google shipped

Google has put an AI meeting assistant on the Mac that does all of its processing on the computer itself. The app is called Google AI Edge Foresight. It is free, Google labels it experimental, and it can transcribe meetings and audio files with no internet connection.11 It runs on macOS and is optimized for Apple Silicon. Google says meeting audio, notes and connected files never leave the user's machine.1113

Reports disagree slightly on the launch date. Several outlets say Google released Foresight on October 6, 2026, alongside a new open model called EmbeddingGemma 2.2718 Others date it to October 8, the day TechCrunch and The Verge covered it.1214 The most likely explanation is that the app and model were announced on the 6th and drew wide attention two days later.

The app works in a simple way. During a call, the user types short bullet points. Foresight listens in the background and turns those fragments into full notes based on what people actually said.211 TechCrunch describes a split screen: the user's shorthand on one side, the AI-generated notes on the other, plus access to the full transcript and a chat assistant for questions afterward.8 The app captures both microphone and system audio, so it works with Google Meet, Zoom or a face-to-face conversation without sending a bot into the call.7

More than a transcriber

The feature that sets Foresight apart from basic note-takers is its knowledge base. Users can add PDFs, Google Docs, Microsoft Office files, plain text, Markdown and web bookmarks. The assistant can then draw on that material, both during a meeting and when reviewing it later. Google also describes a Live Assistance feature that spots questions raised in a meeting and answers them as they come up.918 How-To Geek says the app builds a knowledge graph linking meetings, personal notes and added files.6

The search runs on EmbeddingGemma 2, a 740-million-parameter model that places text, images, audio and video in a single shared vector space. That lets one plain-language question pull up a slide, a transcript excerpt or a document.17 A Gemma 4 model writes the expanded notes and the chat answers on top of what the retriever finds. In other words, this is retrieval-augmented generation, a setup that normally runs in a data centre, running entirely on a laptop.

Here the coverage differs in a way that matters. Some reports suggest EmbeddingGemma 2 does the transcription itself.1 More careful accounts point out that Google has not named its speech-recognition model, and that EmbeddingGemma 2 handles search rather than transcription.7 Hands-on testers found that setup involves choosing a Gemma 4 model and downloading separate "expert models" for speech recognition, question detection and text embedding. That download came to about 800 MB, on top of the optional larger Gemma 4 12B model.3 The more accurate picture is a stack of several small specialized models, not one model doing everything.

The privacy pitch, and its fine print

Privacy is the main selling point. Google's FAQ says all AI inference runs entirely on the device, and it promotes using the app on flights or in air-gapped facilities.7 Analysts see the appeal for healthcare, legal and financial work, where sending meeting audio to a vendor's cloud can stop adoption before it starts.14 One business-focused analysis argues that for many companies, the obstacle to AI in meetings is not output quality but where the recording ends up.9

The word "offline" deserves some scrutiny. One hands-on review found that the setup screen itself lists cases where Foresight connects to the internet: importing from Google Drive and Google Calendar, optional web searches (off by default), and sending de-identified usage data under Google's privacy policy.3 Others note that Google has published no independent audit of its claims. They suggest that anyone handling confidential meetings check the app's network activity themselves.5 The core claim, that audio and transcripts stay on the Mac, still appears solid based on Google's own statements. But "completely offline" describes the AI processing, not every function in the app.

There are also caveats beyond technology. Recording a meeting locally does not remove the need to get participants' consent. Expanded notes are only as reliable as the transcript behind them, and a misheard name or an outdated document can produce a confident but wrong answer.18 One reviewer said the app turned his bullet points into notes quickly. But a single test says little about how it handles overlapping speakers, heavy accents or poor audio, and Google has not published language support or accuracy figures.3

Where it fits among AI meeting assistants

Coverage consistently compares Foresight to Granola. TechCrunch called it a "local-first Granola competitor", and the two apps share the same layout of user notes alongside AI notes.8 Other rivals named include Wispr Flow, Fathom and Notion's AI features.611 Most of these tools send audio or text to the cloud for processing.1 A few smaller products, such as Meetily, already take a local-first approach.1 Foresight did not invent on-device meeting notes, but a Google release gives the idea far more visibility.

The most interesting comparison is with Google's own products. Gemini's "Take notes for me" in Google Meet runs in Google's cloud, saves output as a Google Doc in Drive, and does not join Zoom or Microsoft Teams calls on its own.10 Foresight reverses each of those choices: it works with any platform, stays on the device and keeps files on the user's disk.37 That makes it a natural fit for Mac users who live in Zoom or Teams and do not want their audio leaving the laptop.3

For the wider category of AI productivity and writing tools, the shorthand-to-notes feature is the real change. Foresight is less a transcription engine than a writing assistant that uses the user's own bullets as an outline and the transcript as source material. That design keeps a person deciding what matters, which is arguably healthier than handing over a fully automated summary.

Showcase or product?

The more important question is whether Google intends Foresight to last. TechCrunch notes that the same team released an experimental dictation app called Eloquent earlier this year, which never became mainstream. It suggests Foresight may mainly be a showcase for Gemma's offline capabilities.814 TechCrunch also speculates that Google could release a cloud-based meeting note-taker inside its Gemini apps.217

The app's limitations support the showcase reading. It is Mac-only, Intel Macs are reportedly unsupported, and there is no Windows, iPhone or Android version.7 It is downloaded directly from Google's developer site rather than the Mac App Store.7 Google has published no minimum macOS version or RAM requirement.23 It has not said whether a Windows version or Gmail integration is coming, and it reportedly did not answer follow-up questions.1 One reviewer estimates that a 16GB Mac is the comfortable minimum once Gemma 4 runs alongside the embedding model. That is the reviewer's own estimate, not a Google specification.

The app fits neatly with the developer story. EmbeddingGemma 2 is open under an Apache 2.0 license, and developers can run quantized versions through llama.cpp or Unsloth.1 It also loads in parts: text-only use needs just a 270M encoder.1 Foresight is, in effect, a working reference design: embed everything locally, retrieve by similarity, and generate with a small local model.5

The takeaway

Foresight is best read as a demonstration that also happens to be useful. It shows that transcription, search across a user's documents and writing assistance can now run on a consumer laptop without a subscription or a server.57 That puts pressure on cloud-first note-takers to explain why their users' meetings need to leave the device.

Google has not committed to keeping the app going, and its privacy promises remain the company's own claims rather than independently verified facts.35 For privacy-conscious Mac users, Foresight is worth trying on non-sensitive meetings first.18 For the broader AI meeting assistant market, it suggests that on-device processing is likely to become a standard feature rather than a niche one.