Google AI Edge Foresight Brings Offline AI Meeting Notes to Mac

What Google shipped

Google has released Google AI Edge Foresight, a free and experimental macOS app that records meetings, transcribes them, writes up notes and answers questions about them, all without a network connection.13 The app runs on the device, and Google says it is optimized for Macs with Apple Silicon.2 Google's pitch on privacy is simple: meeting audio, notes and any files you connect stay on your computer.1

The reported launch date varies a little. 9to5Google covered the app on October 6, 2026, the same day Google announced its EmbeddingGemma 2 model, and presented Foresight as a demonstration of that model.12 TechCrunch and The Verge did not run their stories until October 8, which is why some coverage gives that date instead.213 The sequence tells you what Foresight is for. Google built it to show off a model first, and as a product second.

How it works in a meeting

The interface will look familiar to anyone who has used Granola. One side of a split screen holds your own rough notes, and the other holds notes the AI writes.2 You type short bullets during the call, and Foresight uses the transcript to turn them into polished notes.1 Google's product language says this lets you "be completely present" in the meeting instead of typing everything down.15

For audio, Foresight listens to the Mac's microphone and system audio. It does not join the call as a bot.9 That means it works with Google Meet, Zoom and in-person conversations without any platform-specific integration.416 eWeek pointed out how unusual this is for Google: Gemini's note-taking in Meet only works inside Meet, while Foresight doesn't care which meeting service you use.14

Foresight also does more than take notes. It has a built-in assistant you can ask about what was said, and you can connect local files for it to draw on.1 Supported formats reportedly include PDFs, Google Docs, Microsoft Office files, plain text, Markdown and web bookmarks, and notes can be labeled and filtered without a connection.13 A Live Assistance feature is reported to spot questions as people ask them and bring up relevant material from your indexed files.14 Google's product page describes these as context cards with exact citations.15 How-To Geek says the app builds a knowledge graph that links meetings, your notes and the files you add.4

The model question: one engine or two?

Reports disagree about what powers the app. The Verge says Foresight runs on EmbeddingGemma 2.1 TechCrunch adds that the model has 740 million parameters.2 Other outlets say Google credits both EmbeddingGemma 2 and Gemma 4.49 One analysis says the EmbeddingGemma-only version leaves out part of the system.15

The two-model account makes more sense. An embedding model turns text, images, audio and video into a shared numerical space so the app can search and retrieve content. It does not write prose. Coverage describes EmbeddingGemma 2 doing retrieval and Gemma 4 handling the chat assistant. HighTechDad also describes the assistant as an on-device Gemma 4 model.8 One hands-on report counted about 800 MB of required model downloads on first launch, covering speech recognition, question detection and text embedding. The larger Gemma 4 12B download was still pending at that point.16

The Verge and the Windows-focused coverage that followed it also say Foresight transcribes recorded audio files.13 HighTechDad tried importing an audio file and found the current version doesn't seem to support it. Live and in-person recording did work.8 Readers who plan to process a backlog of recordings should check this for themselves before relying on it.

Why local-first matters in a crowded market

AI note-taking is now crowded. Granola and Fathom have built followings, and older tools like Notion have added transcription and summaries.4 Most of these products send audio to remote servers for processing. That includes Google's own Gemini notes in Meet.15 Foresight's main difference is that it skips that step. One breakdown notes that Foresight is free, while Granola charges for business and enterprise tiers.15

This has the biggest impact in regulated work. Several outlets single out legal, medical, HR and finance conversations, where uploading recordings creates compliance problems.913 Google's FAQ also says the app works on flights and inside air-gapped facilities.15 HighTechDad's take is that this is the whole pitch: a Granola-style tool that never sends anything off your laptop, for people who have avoided cloud recording services.8

The analysis here is that Foresight matters less as a product and more as proof that a whole meeting-assistant workflow can now run on an ordinary laptop: capture, transcription, retrieval and chat. Cloud tools have argued that serious language processing needs data-center hardware. A free Google app built on open-weight models makes that argument harder to sustain, at least for one person working alone.

The caveats that coverage agrees on

Coverage agrees broadly on the limits. The first is that Foresight is experimental. TechCrunch notes that the same Google team released a local dictation tool in April that never reached mainstream use, and suggests Foresight may mainly be a showcase for Gemma.2 Other analysts repeat this, warning that experimental Google products can change or disappear without any promise of long-term support.15

The second limit is platform. The app runs only on macOS, and Google has not said whether a Windows version is coming.3 Google has also not said whether Foresight will connect with Gmail.13 How-To Geek expects a Windows port if the experiment works, but that is a guess, not something Google has said.4

The third limit is teamwork. Cloud tools are still better at anything that involves other people. Granola runs on phones and watches, shares notes through team folders and connects to tools like Slack and HubSpot. Foresight is a single-user Mac app with no published sharing or admin features.15 For a company, that means individual employees end up with private transcripts on their own laptops that IT can't see.

The fourth limit is that keeping data local moves the risk rather than removing it. A transcript stored on a laptop is only as safe as that laptop, so disk encryption and screen locks matter more.15 eWeek says organizations need to decide who can access locally stored meeting records, how long to keep them and when to delete them.14 Because the app records through the Mac instead of appearing as a bot, other participants never see a note-taker in the meeting. That is convenient, but it also raises a consent problem.9 Recording laws apply wherever the processing happens, so people need to be told they are being recorded.16

The fifth limit is that privacy and accuracy are Google's own claims. No third-party audit of the privacy promise has been published, and one guide suggests watching the app's network traffic during a test meeting.9 Google has published no accuracy or language figures for transcription.15 Turning shorthand into full notes could link a bullet to the wrong part of the conversation, so the output should be treated as a draft and checked.

The bigger strategic picture

Foresight is part of a wider push behind Google's AI Edge tools. The same LiteRT runtime powers ML Kit, MediaPipe and the AI Edge Gallery demo app on Android and iOS.9 9to5Google reports that Google released Gallery demos alongside the Mac app, with Foresight as the main showcase.12 TechCrunch has speculated that Google could eventually put a cloud-based note-taker in its Gemini apps. One outlet stresses that this is speculation, not a confirmed plan.17

Our reading is that Foresight is most useful as a signal of where things are heading, not as a tool to switch to today. It may never become a supported product. But a free, offline, Granola-like app from the company that trains the models raises expectations for every paid AI meeting assistant. Users will expect sensitive conversations to stay on their own machine, and cloud vendors will need a better reason than "it needs the cloud" to justify a subscription. For now, the sensible approach is the one several reviewers suggest: try it on low-stakes meetings, compare its notes with your own memory of what was said, and wait to see whether Google keeps supporting it before trusting it with confidential work.96