AI Edge Foresight: Google's Simple, Safe Offline Note-Taker

AI Edge Foresight is a new, free and experimental Mac app from Google that listens to your meetings, transcribes them and turns your scribbled shorthand into tidy notes without sending anything to the cloud. Google released it on Tuesday 6 October 2026, alongside a new open model called EmbeddingGemma 2, and The Verge reported that the app transcribes meetings and audio files “entirely offline”.

That makes AI Edge Foresight unusual. Most AI note-takers, from Granola to Google’s own Gemini notes in Meet, record audio on your device and then process it on a company’s servers. Google says Foresight keeps the audio, the transcript, your notes and your files on the Mac, and that the app keeps working on a flight or inside an air-gapped office.

This article explains what Google launched, how AI Edge Foresight works in a real meeting, which Gemma models sit underneath it, and how far its privacy promise really goes. It also compares AI Edge Foresight with Granola and cloud note-takers, and sets out what a business should check before letting staff install it. For the wider shift towards desktop AI on the Mac, see our report on Apple’s warning about AI agents and Full Disk Access.

Table of contents

What Google Launched With AI Edge Foresight
How AI Edge Foresight Works in a Meeting
The Gemma Models Behind AI Edge Foresight
Is AI Edge Foresight Really Offline and Private?
AI Edge Foresight vs Granola and Cloud Note-Takers
AI Edge Foresight Requirements, Languages and Performance
Should Your Business Try AI Edge Foresight?
What Comes Next for Google’s On-Device AI
Frequently Asked Questions About AI Edge Foresight
References and Further Reading

What Google Launched With AI Edge Foresight

AI Edge Foresight is a meeting companion for macOS. Google’s Google Developers Blog post calls it “an experimental app that brings a context-aware meeting companion directly to your device”, powered by “fully local AI processing”. The product page sums it up in one line: “Runs locally on your Mac • Works offline”.

The download is a direct installer from Google rather than a Mac App Store listing. When we checked on 9 October, the disk image on Google’s download server was 160,266,104 bytes, roughly 160 MB, and the server reported it as last modified at 03:04 UTC on 6 October, the morning of the launch.

The AI Edge Foresight launch at a glance

Item	AI Edge Foresight	Source
Released	Tuesday 6 October 2026, labelled experimental	Google Developers Blog
Price	Free, with “no cloud subscription costs”	Google Developers Blog
Platform	macOS, optimised for Apple Silicon; Intel Macs not supported	Product FAQ, heise
Distribution	Direct download from Google, about 160 MB, not the Mac App Store	Google download server, heise
Models	EmbeddingGemma 2 and Gemma 4, run on the Mac	Google Developers Blog
Audio sources	System audio and microphone, so any meeting platform or a face-to-face meeting	Product page and FAQ
Knowledge sources	Local folders and Google Drive; PDFs, Google Docs, Office formats, plain text, Markdown, web bookmarks	Product FAQ
Internet needed	No; Google calls the app “completely offline-first”	Product FAQ
Support	Email to the Google AI Edge support address	Product page

Part of Google’s AI Edge family of apps

AI Edge Foresight joins two earlier apps under Google AI Edge, the banner for Google’s on-device tooling such as LiteRT and MediaPipe. One is AI Edge Gallery, a showcase app for trying on-device models on Android and iOS. The other is AI Edge Eloquent, a dictation app that TechCrunch reported Google quietly released on iOS on Monday 6 April 2026.

According to TechCrunch’s report on Foresight, the same team built both apps. That matters, because Eloquent shows how this team approaches speech. TechCrunch said Eloquent downloads “Gemma-based automatic speech recognition (ASR) models” after installation, then transcribes on the device, with an optional cloud mode for text clean-up that users can switch off.

How AI Edge Foresight Works in a Meeting

The idea behind AI Edge Foresight is that you should be able to pay attention to the people in the room rather than to your notes. Google’s product page puts it as “Be completely present”. You type a few rough bullets as the meeting goes on, and the app fills in the detail afterwards from what was actually said.

Listening to system audio and the microphone

AI Edge Foresight captures sound at the level of the Mac itself. Google’s FAQ says it “runs unobtrusively on your desktop alongside Google Meet, Zoom, or offline conversations, transcribing local system audio in real time”. Because the app hears the computer’s own audio output plus its microphone, it does not need a plug-in for each video platform, and it does not send a bot into the call.

That approach is the same one Granola made popular, which its pricing page describes as “bot-free, device-audio capture”. It also means AI Edge Foresight works for a face-to-face meeting with the laptop open on the table. The trade-off is that nobody else in the meeting sees a visible bot, which raises the consent questions covered later in this article.

Turning shorthand into polished notes

The core feature is what Google calls enhanced note-taking. The product page says: “Simply scribble shorthand bullets for things you want to remember, and Foresight enriches them into polished notes using the meeting transcript instantly.” For example, a bullet such as “budget – Q3 slip?” would become a fuller note built from what people actually said about the budget and the third quarter.

TechCrunch described the layout as a split screen, with your own shorthand on one side and the AI-generated notes on the other, much like Granola. You can also open the full transcript of the meeting. AI Edge Foresight therefore gives you three layers of record: the raw transcript, your own bullets and the expanded notes that join the two.

Chat, Live Assistance and context cards

The second half of AI Edge Foresight is a personal knowledge base. You point the app at “project folders, reference materials, diagrams, and calendar”, and it indexes them without you sorting anything by hand. Sources can carry custom labels and be filtered by category, and Google says you can “semantically search across thousands of pages in milliseconds without cloud indexing”.

You can then ask questions in a chat window. Google says the chat offers “conversational retrieval and summary with your personal knowledge”. A feature called Live Assistance goes further. When a question or topic from your imported files comes up in the conversation, AI Edge Foresight shows “proactive context cards with exact citations”, so you can answer without leaving the call to search your drive.

The Gemma Models Behind AI Edge Foresight

Google describes AI Edge Foresight as “powered by EmbeddingGemma 2 and Gemma 4 models”. Some reports, including The Verge’s, described the app as running on EmbeddingGemma 2 alone. That is incomplete, and the difference matters if you want to understand what the app can and cannot do.

EmbeddingGemma 2 finds things; it does not transcribe

EmbeddingGemma 2 is an embedding model. As Notebookcheck explained, “Embedding models do not generate text. Instead, they convert content into numerical vectors so that similar material can be retrieved.” The new model places text, code, images, video frames and audio in one shared vector space, so a typed question can find a matching paragraph, slide or stretch of a recording.

In AI Edge Foresight, that is what powers search, the knowledge base and Live Assistance. Google’s EmbeddingGemma 2 launch post describes this pairing directly: “Pair EmbeddingGemma 2 for local file retrieval with Gemma 4 for contextual reasoning. Try it in the Google AI Edge Foresight app.”

Gemma 4 does the writing, and probably the listening

Writing polished notes, answering chat questions and summarising a meeting all need a generative model, and that is Gemma 4. TechCrunch referred to “the Gemma 4-powered assistant” in AI Edge Foresight. Google’s DeepMind page lists Gemma 4 in E2B, E4B, 12B, 26B and 31B sizes, and says the two smallest bring audio and vision support to edge devices. Google has not said which size the app runs.

Google has also not named the speech-recognition model that produces the transcript. The FAQ says only that “All AI inference runs 100% locally on your device using on-device Gemma models.” Two clues point in the same direction. EmbeddingGemma 2 “shares its text tokenizer and audio encoder” with Gemma 4, and the team’s earlier Eloquent app uses Gemma-based speech recognition. A Gemma-family speech model is the most likely answer, but it is an inference, not a confirmed fact.

EmbeddingGemma 2 by the numbers

Specification	EmbeddingGemma 2
Parameters	740 million: 270M text, 170M vision encoder, 300M audio encoder
Architecture and licence	Built on Gemma 4; Apache 2.0
Inputs	Text, code, images, video frames and audio in one vector space
Context window	8K tokens: up to 5.5 minutes of audio, 29 images or 58 video frames
Output size	768 dimensions, truncatable to 512, 256 or 128
Memory on a Pixel 11 Pro	About 191MB text-only; about 567MB with all encoders (quantised)
Image speed	37.3 ms per image on a MacBook M5 Pro GPU, about 26.9 images a second
Code retrieval (MTEB Code)	78.68, up from 68.76 for the first EmbeddingGemma
Where to get it	Hugging Face, Kaggle and LiteRT Community bundles

The modular design is why the model suits a laptop app. Developers can load only the encoders they need, so an app that only searches documents never pays for the audio encoder. The chart below splits the 740 million parameters by component, with each bar showing that component’s share of the total.

EmbeddingGemma 2 parameters by component (share of 740M)

Audio encoder, 300M 40.5%

Text model, 270M 36.5%

Vision encoder, 170M 23.0%

The figures come from Google DeepMind’s launch post, and each percentage is the component’s parameter count divided by 740 million. The audio encoder is the largest single part, which fits a model meant to search recordings. It also explains the gap between the two memory figures: the full multimodal model needs roughly three times the RAM of the text-only version.

Is AI Edge Foresight Really Offline and Private?

Privacy is the reason AI Edge Foresight exists, so its claims deserve a careful read. The short answer is that Google makes strong, specific promises about processing, but the app is experimental, its terms are generic, and a local app does not remove your own legal duties when you record people.

What Google promises

Google’s FAQ is unusually direct. It says “Your personal knowledge sources, meeting audio, and notes never leave your computer” and that “No personal files, recordings, or transcripts are ever transmitted to external cloud servers”. It adds that “Both the Gemma model and the local vector embeddings engine run on your hardware without internet”, so you can “travel, fly on airplanes, or work in air-gapped secure facilities”.

For anyone who has avoided cloud note-takers because of client confidentiality, that is the selling point. A cloud service holds your transcripts on someone else’s servers, under someone else’s retention rules. With AI Edge Foresight, the transcript is a file on your own disk. The heise report contrasted this with cloud-based Mac assistants such as Meta’s Muse, which shipped with a serious security flaw.

Where the promise has edges

Local processing moves the risk rather than removing it. A transcript that lives on a laptop is only as safe as that laptop, so disk encryption, screen locks and backups now carry the weight that a vendor’s controls carried before. From a cybersecurity point of view, a stolen, unencrypted Mac full of meeting transcripts is a bigger problem than one with a few documents on it.

There are smaller gaps too. AI Edge Foresight can connect Google Drive, and a Drive file has to come from Google’s servers before it can be indexed locally, so “works offline” describes processing, not every source. The product page links only to Google’s general privacy policy and terms of service, not to a product-specific notice. And heise noted that one should still be cautious about what information is shared with the app, especially as Apple moves to restrict the broad file access that many AI tools ask for.

Recording law and consent still apply

Running the model locally does not change the law on recording people. Under UK data protection law, a meeting recording and its transcript are personal data, so a business needs a lawful basis and must tell people what is happening. The Information Commissioner’s Office guidance on the right to be informed applies whether the audio goes to a cloud or stays on a Mac.

Because AI Edge Foresight captures system audio without a visible bot, nobody else in the call gets an automatic notice. Granola’s Enterprise plan, by contrast, can show an “org-wide notification that Granola is being used”. With Foresight, telling participants is entirely up to the person running the app, which is why a written policy should come before any rollout.

AI Edge Foresight vs Granola and Cloud Note-Takers

TechCrunch called AI Edge Foresight a “local-first Granola competitor”, and the comparison is fair on features. The bigger difference is where the work happens. CNET noted that most popular note-takers, Granola among them, use the cloud to access and store data, and Google’s own Meet notes run on Gemini in Google’s cloud. Foresight processes everything on the Mac, which changes the cost, the privacy picture and what the app can share.

Feature comparison

Feature	AI Edge Foresight	Granola	Google Meet “Take notes for me”
Where audio is processed	On the Mac	Cloud, per CNET	Google’s cloud (Gemini)
Price	Free, experimental	Free Basic; $14 Business; $35 Enterprise (per user per month)	Google AI Pro or Ultra for consumers, or eligible Workspace editions
Platforms	Mac with Apple Silicon	Desktop, iOS, Android and Apple Watch	Meet on web, Android and iOS
Capture method	System audio and microphone	Bot-free device audio	Inside Meet, or a phone nearby for in-person meetings
Where notes end up	On the Mac	Granola workspace and shared folders	A Google Doc in Drive, with a link emailed
What you get	Transcript, enriched notes, chat and Live Assistance from your files	AI notes and AI chat within and across meetings	Transcript, summary and action items
Works without internet	Yes	Not advertised	No, it is a cloud service
Admin controls	None published	SSO, admin controls and org-wide auto-deletion on Enterprise	Workspace admin settings

Granola’s prices and features come from its pricing page on 9 October 2026. The Meet details come from 9to5Google’s August report on in-person note-taking, which said the feature saves “a complete transcript, summary, and list of action items” to a Google Doc in Drive.

What a ten-person team would pay in a year

The clearest practical difference is cost. AI Edge Foresight has no subscription at all. The chart below works out a year of Granola for a ten-person team at its published per-user monthly prices: $14 × 10 users × 12 months for Business, and $35 × 10 × 12 for Enterprise. Bars are scaled to the Enterprise total.

Annual note-taker cost for a ten-person team (US dollars)

Granola Enterprise, $35 per user per month $4,200

Granola Business, $14 per user per month $1,680

Granola Basic, 30 days of history $0

Google AI Edge Foresight, experimental $0

The zero-cost bars are drawn at a sliver so they stay visible. A free experimental app is not the same as a free product, though. Foresight’s real costs are the staff time to set it up, the Mac hardware it needs and the risk that Google withdraws or changes it. Granola Basic is also free, but it keeps only 30 days of meeting history.

Where cloud note-takers still win

Cloud tools are better at anything that involves other people. Granola runs on phones and on the Apple Watch as well as desktops, shares notes through team folders, and plugs into Notion, HubSpot, Slack and Zapier. AI Edge Foresight is a single-user Mac app, and Google has published nothing about sharing, team workspaces or integrations beyond its file sources.

The market is also moving quickly. TechCrunch noted that Wispr and Calendly have both moved into meeting notes, and that Superhuman (formerly Grammarly) bought the note-taker Fathom in September. Bots are spreading too: we covered xAI’s work on a Grok meetings feature that joins Google Meet and Zoom, and Agora’s open-source meeting copilot built on GPT-Live-1. Against that crowd, AI Edge Foresight stands out mainly because it runs offline.

AI Edge Foresight Requirements, Languages and Performance

Google has published far fewer details about the app than about the model, so several practical questions are still open. Here is what is confirmed, what is measured on other hardware, and what is still unknown.

Hardware and installation

AI Edge Foresight is “optimized for Apple Silicon”, and heise reported that Intel Macs are not supported. Google has not published a minimum macOS version, a minimum RAM figure or a disk-space requirement for the models. It says only that “lightweight Gemma quantization and memory-efficient vector kernels” give the app “a light footprint with low battery drain”.

CNET reported that the app is “only on Mac for now” and that it is not known whether or when a Windows version will follow. There is no iPhone version of AI Edge Foresight either. The closest Google offers on mobile are the new EmbeddingGemma 2 search demos in AI Edge Gallery, and Eloquent for dictation.

Languages

EmbeddingGemma 2 supports more than 100 languages for search, according to heise. Transcription is a separate question. Notebookcheck noted that Google “has not said whether transcription works equally well for meetings held in languages other than English”. The team’s Eloquent app, which shares this speech heritage, launched in English only, and heise reported it has since added German. Anyone holding meetings in Welsh, Polish or Hindi should test before relying on the transcripts.

Speed and storage figures

Google’s published speed and memory figures are for the model, not for the app in a live meeting. The 37.3 ms image figure was measured on a MacBook M5 Pro GPU, and the RAM figures on a Pixel 11 Pro phone. Treat them as a sign of scale rather than a guarantee for your own Mac.

The storage figures are easier to reason about, because they follow directly from the vector size. The table below assumes each dimension is stored as a standard 4-byte float and that a knowledge base holds 10,000 indexed chunks.

Vector size	Bytes per vector	10,000 chunks	Saving vs 768
768 dimensions	3,072	30.7 MB	None
512 dimensions	2,048	20.5 MB	1.5 times smaller
256 dimensions	1,024	10.2 MB	3 times smaller
128 dimensions	512	5.1 MB	6 times smaller

The six-times saving at 128 dimensions matches Google’s “up to 6x storage reduction” claim. Audio follows a similar logic: at 5.5 minutes of audio per 8K-token input, a 60-minute recording would need at least 11 separate inputs if it were embedded as raw sound. Google has not said how AI Edge Foresight splits recordings, so this is arithmetic on the model’s limits rather than a description of the app.

Should Your Business Try AI Edge Foresight?

For most organisations, AI Edge Foresight is worth a small, controlled test rather than a rollout. It is free, its privacy design is attractive, and its features match the paid leaders. But it is experimental, Mac-only and has no admin controls, so it suits individuals and pilots better than whole departments.

Where AI Edge Foresight fits well

The strongest case is sensitive meetings that should not be transcribed in the cloud: legal and HR discussions, board meetings, client calls under strict confidentiality terms, and early product or deal talks. It also suits people who work on trains and flights, and teams in secure sites with no internet at all, which Google names directly in its FAQ.

Teams already on Macs with Apple Silicon have the least to lose by trying it. Because the knowledge base is local, an individual can point AI Edge Foresight at their own project folders and get useful answers without any IT integration work.

Reasons to wait

Experimental Google products can change or disappear, and Google has not promised long-term support. TechCrunch suggested Foresight may be mainly a showcase for the Gemma models, noting that the team’s earlier dictation app “didn’t make it to mainstream usage”. There is also no published admin console, no central logging and no way to enforce a notification policy across staff.

Mixed estates are a further problem. If half your staff use Windows or Intel Macs, an app that runs only on Apple Silicon creates two classes of meeting record. And without speaker labels in Google’s published feature list, transcripts may be harder to use as formal minutes than those from tools that tag each speaker.

A pilot checklist

Step	Why it matters
Write a short recording policy first	The app gives no automatic notice to other participants
Confirm FileVault and screen locks on pilot Macs	Transcripts and indexes live on the laptop, not with a vendor
Decide where notes are kept and for how long	Local files fall outside any cloud retention rules you already run
Test transcription in your accents and languages	Google has not published accuracy or language figures for the app
Limit which folders the app may index	Answers can only be as safe as the sources you connect
Watch for a data protection impact assessment trigger	New monitoring or recording tools can count as high-risk processing under UK GDPR

The ICO’s pages on data protection impact assessments and monitoring workers are the right starting points. If you already use cloud transcription such as Gemini 3.5 Transcribe, the same policy can cover both, with a note on which tool to use for which kind of meeting.

What Comes Next for Google's On-Device AI

AI Edge Foresight is one part of a bigger push. Google says EmbeddingGemma 2 will come to Android “as a service” through ML Kit “in the coming weeks”, with NPU acceleration where phones support it. MediaPipe is adding the model to its Embedder and Semantic Retriever tasks for iOS, macOS, Windows, Linux and the web, so developers can build similar local search into their own apps.

That points to two possible futures for the app. Google could grow AI Edge Foresight into a supported product, perhaps on Windows, now that the underlying tools run there. Or, as TechCrunch suggested, it could stay a showcase while Google brings a cloud note-taker to its Gemini apps. Either way, the launch shows that an offline AI note-taker no longer means giving up chat, search or summaries.

Frequently Asked Questions About AI Edge Foresight

Is AI Edge Foresight free?

Yes. Google offers AI Edge Foresight as a free download and says it has no cloud subscription costs. It is labelled experimental, so Google could change or withdraw it.

Does AI Edge Foresight work on Windows, Intel Macs or iPhone?

No. AI Edge Foresight is a macOS app optimised for Apple Silicon, and heise reported that Intel Macs are not supported. CNET reported there is no word on a Windows version, and there is no iPhone version.

Does AI Edge Foresight need an internet connection?

No. Google says both the Gemma model and the local embeddings engine run on your hardware without internet. You only need a connection to download the app and to reach online sources such as Google Drive.

Which model transcribes the audio?

Google has not named it. The app is “powered by EmbeddingGemma 2 and Gemma 4”, and the FAQ says all inference uses on-device Gemma models. EmbeddingGemma 2 handles search rather than transcription.