Google AI Edge Foresight Brings Offline AI Meeting Notes to Mac

Google has launched AI Edge Foresight, an experimental Mac app that tackles a familiar problem: getting useful meeting notes without handing every conversation to a cloud service.

Foresight can listen to system audio and microphone input, create transcripts, expand brief notes into more complete summaries and answer questions using documents stored on the computer. Google says the processing happens on-device and the app can keep working without an internet connection.

That makes it an interesting alternative to the growing number of AI note-takers built around uploading recordings for transcription. But there is an important limitation: the initial release is designed for Apple Silicon Macs, not Windows PCs or Android phones.

What Foresight actually does

Instead of asking users to remember every important statement during a call, the app lets them jot down short observations while it follows the conversation. It can then enrich those notes with details from the transcript.

For example, someone could write that the team agreed on a Friday deadline. Foresight can use the recorded discussion to provide more context about the decision and any related action items. The point is to make a person's own notes more useful rather than generate a disconnected summary from scratch.

Google also describes real-time question answering during meetings. If somebody mentions a figure or document, Foresight can use information indexed locally to help retrieve related context.

Because it captures both system audio and microphone input, Google says it works across meeting platforms instead of requiring an individual Zoom or Google Meet integration. It may also be used for in-person discussions where the microphone can hear participants.

See Google's official Foresight page.

Two Gemma models handle different jobs

The app uses EmbeddingGemma 2 and Gemma 4, according to Google's developer announcement.

EmbeddingGemma 2 is a compact, approximately 740-million-parameter multimodal embedding model. Rather than writing a meeting summary itself, an embedding model converts different inputs into representations that can be compared to help search for relevant material. That can make it possible to find the right note, transcript passage, image or document using an ordinary-language query.

Gemma 4 helps power the assistant side of the experience, including answering questions and producing readable notes from retrieved context. These are separate responsibilities, and the distinction is important: searching for the right information and composing a response are not necessarily the same model operation.

Google is also using EmbeddingGemma 2 in mobile demonstrations for searching images and locating particular moments in videos. Foresight is a more practical productivity showcase for the same local AI approach.

Read Google's EmbeddingGemma 2 developer announcement.

Why offline processing matters

Meeting recordings can contain personal details, business plans, client information and discussions that should not become public. A local-first workflow can reduce the need to transfer that audio and associated documents to a remote service for inference.

It also helps in situations where a connection is unstable, such as travel, field work or an office with restricted internet access. And Google says Foresight does not require a cloud subscription to perform its core local tasks.

Local processing is not the same as automatic security, however. Users still need to protect their devices and backups, check where recordings are stored, and get consent where recording or transcription rules require it. Workplaces may also have their own policies on meeting capture and confidential data.

Can Windows users try it?

Not this Mac app, at least for now. Google's announcement identifies Foresight as a Mac release, and coverage of the launch says it is optimized for Apple Silicon. That matters for PC owners interested in local models on Windows, because compatibility should not be inferred just because Google's underlying AI technology can operate across different platforms.

Developers can experiment separately with Google AI Edge and EmbeddingGemma 2 using the company's supported libraries and demos. That is not the same thing as an official Foresight-for-Windows application.

For phone users, Google AI Edge Gallery offers other local demonstrations involving searches through photos and video moments. Those are distinct experiences rather than a mobile version of the Foresight meeting app.

How is it different from a conventional meeting recorder?

The key difference is its combination of local audio capture, interactive note expansion and retrieval across private documents. A simple recorder only collects audio. A transcription tool creates text. An AI assistant that also indexes relevant local files may help connect the discussion with the material people were working from.

That can be especially useful for researchers, students, product teams and people who attend long meetings. It may also appeal to businesses cautious about uploading sensitive calls to third-party servers, although organizations should verify the app's security posture and usage policies before deploying an experimental tool at scale.

The other distinction is maturity. Google calls Foresight an experimental application, so it should not yet be treated as a fully established business-records system. Users should verify transcripts, confirm action items and keep an independent copy of any critical meeting decisions.

What to watch next

The interesting question is whether the underlying local AI features will make their way into broader Google products. Google is already working to expose EmbeddingGemma 2 through Android ML Kit, which could support new device-side search experiences. That future developer availability does not confirm Foresight will come to Windows, Android or Google Workspace.

For now, Foresight is a timely example of what running smaller, specialized AI models on a personal computer can accomplish. Rather than compete solely on enormous model size, Google is demonstrating a specific job that local AI can perform with a clear privacy and connectivity benefit.

For users with a supported Mac, the most sensible next step is to consult Google's official app page and test it with non-sensitive material before trusting it with important meetings.