Google AI Edge Foresight: Free Offline Meeting Notes on Mac

Ever finished a meeting with a few rushed bullets and wished you could fill in the gaps without uploading the conversation? Google AI Edge Foresight is a free, experimental Mac app that records and transcribes meetings locally, then uses Gemma 4 and EmbeddingGemma 2 to help turn your shorthand into fuller notes. Released on October 6, 2026, it captures your microphone and computer audio, so you can use it across meeting apps or in person.

Once you have downloaded its models, you can record and use its AI without an internet connection. Here are the key details before you try it.

Detail	What we know
Developer	Google’s AI Edge team
Released	October 6, 2026
Price	Currently free and experimental
Platform	Macs with Apple Silicon
AI models	Gemma 4 plus the 740M-parameter EmbeddingGemma 2
Internet needed	Initial setup requires it, offline use follows
Works with	Zoom, Google Meet, other meeting apps and in-person conversations
Download	Google’s Foresight download page

What Google Built and Why

You might see Foresight as another way to get meeting notes, but Google built it to demonstrate what EmbeddingGemma 2 can do on your own hardware. In its Developers Blog announcement, Google describes an experimental meeting assistant that uses the discussion and your documents to provide relevant help. You get a practical example of local AI handling information across documents, images and meeting records. To understand where it fits, here are the three related AI Edge apps.

• AI Edge Gallery. You can try Google’s on-device AI demonstrations, including tools for searching media and finding particular moments within local videos. Those media demonstrations are available in Gallery on Android and iOS.
• AI Edge Eloquent. You get local speech transcription and dictation. TechCrunch reports that the same Google team released an experimental dictation app in April 2026, several months before Foresight arrived.
• AI Edge Foresight. You get meeting transcription, help expanding your notes and answers drawn from material you have connected to the app. Its meeting tools run on your Mac.

You see, the meeting itself is only part of what Foresight works with. Your own files supply additional context, which brings us to what you can actually do during and after a call.

Main Features of Foresight

You can use Foresight to write meeting notes, bring relevant documents into the discussion and ask questions afterward. Google says the processing happens on your Mac, with different models handling retrieval and assistant tasks. Here is how those features fit into your meeting.

Live Notes From Your Shorthand

You get two panels, with your quick notes on one side and the expanded version on the other. As you type bullets, Foresight uses the ongoing transcript to add the surrounding detail.

You can still write your own notes while transcription continues. That means you choose what to jot down, while the app uses what was said to help fill out those points.

A Private Knowledge Base

Your meeting may refer to information that nobody reads aloud. You can connect local folders or Google Drive, with support for PDFs, Google Docs, Microsoft Office documents, plain text, Markdown and saved web links.

You can connect your calendar too, giving the app additional meeting context. If you have explored AI tools that act as a second brain, the idea will feel familiar. Your saved material becomes something you can ask questions about, including during a discussion.

Live Assistance and Chat

What happens when someone asks a question your documents already answer? Live Assistance can recognize questions in the conversation and retrieve relevant information from your connected files.

After the call, you can read the transcript or ask the Gemma 4 assistant questions about the meeting. You can search earlier meeting records too, so your questions do not have to stop at the conversation you just recorded.

The Models Under the Hood

EmbeddingGemma 2 handles finding relevant material, while Gemma 4 powers the assistant. You can see the main EmbeddingGemma 2 specifications below, though the memory and speed figures come from particular devices rather than a published Foresight minimum requirement.

Specification	Reported detail
Parameter count	740 million
License	Apache 2.0
Active RAM on Pixel 11 Pro	Around 191MB for text weights or 567MB for the complete multimodal model
Image processing on M5 Pro	About 37.3 milliseconds per image using the MacBook’s GPU
Vector output	768 dimensions, with smaller representations available to reduce storage and index size by up to eightfold

What You Need to Run It

Before you download Google AI Edge Foresight, check whether your Mac has Apple Silicon. Google has not published a minimum chip generation or macOS version so far, so suggestions such as M1 or newer remain third-party guidance. If you are shopping for a new machine anyway, our look at the M6 MacBook Pro covers what Apple has coming. Here is what you can check before getting started.

Requirement	Status	Source note
Mac processor	You need Apple Silicon. Intel Macs are unsupported.	Reported by By Burk. TechCrunch cites Google’s FAQ for Apple Silicon optimization.
Minimum chip	No official minimum identified. M1 or newer is an author’s recommendation.	Reported guidance from Medium, not a confirmed Google requirement.
macOS version	No official minimum identified.	Unconfirmed in the available sources.
App download	About 153MB, expanding to 376MB after extraction.	Reported by talat.
Model downloads	Roughly 800MB for speech and search, plus a 3.41GB language model.	Reported by talat for its tested setup.
Google account	You can start without one. Sign-in is needed for optional Drive and Calendar connections.	Reported by talat.
Internet access	You need a connection to download the models before using the app offline.	Setup reported by talat. Offline operation confirmed by Google.

How to Install and Start Your First Meeting

You will need to download the app and its models before recording your first conversation. The steps below draw on a third-party setup guide and talat’s early test.

1. Check your processor. Open About This Mac and confirm that your computer uses Apple Silicon. Intel machines are unsupported, according to By Burk.
2. Download Foresight. Get the installer from Google’s official download page. You download it directly, rather than through the Mac App Store.
3. Move it into Applications. Open the app from there. The setup guide suggests right-clicking and choosing Open if macOS displays a warning.
4. Download the models. Allow the initial speech, search and language downloads to finish. In talat’s test, setup took roughly eight and a half minutes on a fast connection.
5. Approve audio access. Give the app permission to use your microphone and capture computer audio. Restart it if you are asked.
6. Connect your Google account if needed. Sign in if you want Drive or Calendar access. You can begin without connecting either service.
7. Add a folder. Start with documents you know well, such as a few PDFs. This gives you material against which to check retrieved answers.
8. Start recording yourself. Get everyone’s consent before recording. Press the recording control manually, since talat found no automatic start feature.
9. Write a few bullets. Type shorthand while the conversation continues and check the expanded notes. If nothing happens, check whether Live Enhance is enabled, as talat needed to switch it on.

If your first test stalls, the setup guide suggests these checks.

• Missing transcript. Check both audio permissions and your selected audio output.
• Slow file indexing. Begin with text files before adding other material.

How Private Is It Really

You can record and use Foresight’s AI offline after setup, according to Google, which says your meeting content stays on the Mac. However, talat’s early test found outgoing usage analytics, so offline capability does not mean the app never connects to the internet. Here is what Google claims and what outside testers observed.

What Stays on Your Mac

• Meeting content. Google says your recordings, transcripts, notes and connected files remain on your computer while its models process them locally.
• Observed audio handling. During its brief test, talat did not detect recordings or transcript content leaving the Mac. That supports Google’s claim within the limits of that test.
• Offline AI processing. Google says you can use the app without cloud processing or a cloud subscription. The initial model downloads still require internet access, according to talat.

What Still Goes Online

• Usage analytics. Talat recorded 74 analytics reports in roughly two minutes, covering details such as meeting duration, word totals and model activity. It found no transcript text in those reports.
• Installation identifier. Talat observed an anonymous installation ID and a setup disclosure about de-identified usage information. The tester could not find a way to disable collection.
• Background activity. Talat reported that a helper remained active after quitting, with a scheduled task reopening the app every four hours to update calendar events.
• Google connections. By Burk questioned how optional Drive access fits Google’s local-data claims. Its report raised the question without establishing that meeting recordings were uploaded.

Foresight vs Granola Compared

If you are choosing between Foresight and Granola, start with where you want your meeting data processed. Foresight runs its AI on your Mac, while Granola relies on cloud services for transcription and summaries. That difference affects offline use, but your devices and team requirements matter too, as the comparison below shows.

Feature	Google AI Edge Foresight	Granola
Price	Currently free during its experimental release	Free Basic plan, with Business at $14 and Enterprise at $35 per user monthly
Where AI runs	Locally on your Mac using Gemma models	Cloud services, including Deepgram, AssemblyAI, OpenAI and Anthropic
Offline	Recording and AI work without internet after setup	Transcription and summaries require a connection
Platforms	Apple Silicon Macs	Mac, Windows, iOS, Android and Apple Watch
Bot-free capture	Records your microphone and computer audio without joining the call	Captures device audio without adding a meeting bot
Speaker labels	Separates your speech from other voices, without individual names in talat’s test	Supports speaker tagging
Knowledge base	Retrieves answers from local documents and connected Drive files	Lets you ask questions across meeting records
Integrations	Google Drive and Calendar	Slack, Notion, Zapier, HubSpot and Salesforce, depending on your plan
Maturity	An experimental release with limitations reported in early testing	An established commercial service with paid team plans

If you work alone on a supported Mac and want local processing, Foresight gives you a free option to test. If confidentiality drives your choice, check the analytics and account connections before using sensitive material. Granola is the stronger fit when you need Windows or mobile access, speaker tagging and connections to team software.

You see, these tools serve different working habits. Before you commit to Foresight, its early limitations deserve a closer look.

Known Limits Before You Rely on It

Your first few meetings may expose rough edges, judging by early hands-on reports. Most of the issues below come from talat, a competing note-taking app whose tester used Foresight briefly, so treat them as initial findings.

• An uncertain future. By Burk notes that the experimental app could be withdrawn. You should not assume it will become a permanently supported product.
• Manual recording. In talat’s test, you had to start each recording yourself. Joining a meeting did not trigger it.
• Limited speaker labels. You could distinguish your speech from the other participants collectively, but individual speakers were not identified.
• Speaker echo. Talat found that audio played through speakers could be captured twice, producing repeated transcript content.
• Uneven language results. Google has not published a transcription language list. Talat reported strong English results, missing accents in French and a failure to follow Arabic.
• Restricted model choice. You cannot substitute your preferred AI model, according to talat. The available choices use Gemma.
• No Windows release. You need a supported Mac, and a Windows version has not been announced.
• Answers need checking. Talat reported a weak mid-meeting summary and another summary that introduced a discussion that had not happened.

What This Release Says About Google’s Plans

You can read Foresight as a practical demonstration of Google’s offline models, which is how TechCrunch interprets the release. Its current value is something you can test today, while any larger product strategy remains uncertain. That distinction matters when considering what might come next.

TechCrunch speculates that Google could eventually offer a consumer meeting assistant through Gemini using cloud models. You should treat that as the publication’s prediction, not an announced product or development plan.

You already have a growing set of meeting assistants to choose from. TechCrunch reports recent launches from Wispr and Calendly, plus Superhuman’s acquisition of Fathom the previous month. For now, Foresight gives you another option built around processing on your own computer, the same idea behind the push for an AI PC. Google’s cloud side keeps moving too, as our Gemini 4 Argon guide shows.

Should You Try Foresight Today

If you have a supported Mac, you can give Foresight a small trial before making it part of your working day. Let your first test answer whether it fits the meetings you actually have.

Try It If

• You want free meeting notes with AI processing on your Mac.
• You need recording and AI access without internet after setup.
• You are comfortable checking transcripts and generated answers.

Wait If

• You need Windows, mobile access or individual speaker names.
• You depend on automatic recording or established team integrations.
• You need clearer answers about analytics before handling confidential conversations.

Start with a short, non-sensitive meeting and get everyone’s consent before recording. Compare the notes with what was said, then test a few questions against familiar documents. If the results hold up, try it in another meeting and build from there.

FAQs