Google has released AI Edge Foresight, an experimental Mac app that records and transcribes meetings on the device and turns a user's shorthand bullets into fuller notes. Google's developer blog announced it on 6 October 2026 as the showcase for a new open model, EmbeddingGemma 2. For anyone choosing an AI note taker, the interesting part is not the app itself but the design choice: no cloud processing, and no cloud subscription cost.
What Google announced
In its post, Google describes Foresight as an experimental app that brings a context-aware meeting companion to your device. The blog says it works with the system audio and microphone, so it should work with any meeting platform, and that it can run completely offline. Google lists three things the app delivers: notes enriched from the live transcript, search across different media types using plain language, and on-device processing with no cloud subscription costs.
9to5Google reports that the app listens in the background while you scribble shorthand bullets, then expands them into polished notes using the transcript. It says the transcript and audio are processed entirely on your device rather than in the cloud. Users can also point the app at project folders, reference material and a calendar, with support for local files and Google Drive, and ask it questions. A Live Assistance feature listens for questions in the meeting and answers them.
TechCrunch describes the app as a competitor to the note taker Granola, with a split-screen view for your own notes and the AI's notes. It adds that you can read the transcript or chat with an assistant powered by Gemma 4, and that the app can capture in-person meetings as well as calls. It lists PDFs, Google Docs, Microsoft Office files, plain text, Markdown and web bookmarks as material you can add to a knowledge base.
The models behind it
Two Google models do the work. Gemma 4 powers the assistant, and EmbeddingGemma 2 handles search. According to 9to5Google, EmbeddingGemma 2 has 740 million parameters, is built on the Gemma 4 architecture and carries an Apache 2.0 licence. Google's blog says that, with quantization, it needs as little as about 191MB of active memory for text-only weights and about 567MB for the full multimodal model, measured on a Pixel 11 Pro. Those figures are for the search model on a phone, not for the whole Mac app, and none of the sources we read gives a footprint for Gemma 4 or a minimum Mac specification.
The same post says the Android version of the model will arrive through ML Kit in the coming weeks. Tech AI Wire's summary of the coverage says TechCrunch cites Google's FAQ for the app being Mac only and optimised for Apple Silicon. If you run Windows or Linux, or want a note taker on a phone, this release is not for you yet.
Why it matters to buyers
Most popular note takers send meeting audio to a vendor's servers, where it is transcribed and summarised. That raises questions buyers should ask of any product: where the audio is stored, who can access it, and whether it is used to train models. A tool that keeps everything on the laptop changes those answers, at least for the recording and transcription step. Google says sensitive data stays on the device.
There are limits to read into that claim:
- It is experimental. Google labels it so. Do not treat it as the only record of an important meeting.
- It is a showcase. TechCrunch suggests the app mainly demonstrates Google's offline Gemma models, per Tech AI Wire's summary. A showcase app may not receive the support, integrations or admin controls a paid team product offers.
- Local is not the same as consent. Recording a call on your own machine still captures other people. Rules on recording and on telling participants vary by place and by workplace policy, so check them, and ask a qualified professional if you are unsure.
- Cloud features you may rely on are not mentioned. The sources do not describe team sharing, CRM sync or shared workspaces, which paid tools typically sell. Do not assume they exist.
The wider market
The launch lands in a busy month for the category. Tech AI Wire, summarising TechCrunch, notes that Wispr and Calendly have released their own note takers in recent months and that Superhuman bought Fathom in September. Fathom is one of the tools in our AI note taker category, so existing users may want to watch for any change to its terms or data handling under new ownership. We have not verified any such change.
What to do
- Decide what you actually need. If you only want private notes from your own meetings on a recent Mac, an on-device tool is now a serious option to evaluate. If you need shared notes, integrations or support, a cloud product still covers more ground.
- Read the vendor's data page before you rely on any note taker. Look for where audio is stored, how long it is kept and whether it trains models. For Foresight, use Google's Foresight page as the starting point.
- Tell participants. Whether the tool is local or cloud, say that a meeting is being recorded and follow your workplace policy.
- Try before you move. If you pay for a cloud note taker, running an experimental tool alongside it on a low-stakes meeting costs little and shows whether the notes are good enough for you. Keep your existing tool until you are sure.
- Keep expectations modest. Experimental software can change or disappear, and the sources do not say how long Google will support it.
The release does not make cloud note takers obsolete, and it does not prove local notes are better. It does show that a capable meeting assistant can run on a laptop, which gives buyers a stronger question to put to vendors: what exactly needs to leave my device, and why?
