TechCrunch reported on October 8 that Google has released Google AI Edge Foresight, an AI meeting note-taker for the Mac built around a "local-first" promise: all transcription and AI processing happen on the device, no internet connection required. The launch positions Google directly against Granola, the red-hot AI note-taking app that has become the reference point for the category.
Foresight borrows Granola's signature split-screen workflow. During a meeting you scribble shorthand bullet points on one side while the app enriches them into polished notes on the other, drawing on the live meeting transcript. It listens through the microphone and system audio, so it works with any meeting platform — and with in-person conversations too — with no need to invite an AI bot into the call. Questions raised during the meeting are detected automatically and answered against your personal knowledge base, and you can pull up the full transcript or chat with the built-in Gemma 4 assistant at any time.
According to Google's developer blog, all transcript and audio processing takes place on the device, with no cloud subscription required. Under the hood is EmbeddingGemma 2, released the same day by Google DeepMind: a 740-million-parameter multimodal embedding model built on the Gemma 4 architecture and released under the Apache 2.0 license. With quantization it needs only about 191MB of active RAM for text-only weights — about 567MB for the full multimodal model — on a Pixel 11 Pro. Its 8,192-token context window is four times the previous generation's, good for roughly 5.5 minutes of audio, 29 images, or 58 video frames, and Google says image processing takes just 37.3 milliseconds on an M5 Pro GPU. As Notebookcheck put it, on-device AI that keeps all data local was supposed to be Apple's promise; Google is now building a Mac app around the same claim.
Foresight lets you upload PDFs, Google Docs, Microsoft Office files, plain text, Markdown documents, and web bookmarks to build a personal knowledge base, and it can also be pointed at project folders, diagrams, and calendars. When a discussion touches on something in that knowledge base, the app surfaces relevant answers in real time. 9to5Google reports the pitch as an app that understands your work without you having to organize files manually. TechCrunch notes the team previously shipped a local AI dictation tool in April — an experiment that never broke through — and has now aimed the same technology at a stickier use case.
Granola is the name to beat: it raised $125 million in March 2026 on the thesis that transcripts are the richest source of context inside a company. But the field is crowded. Dictation company Wispr and scheduling company Calendly both launched their own notetakers in recent months, and on October 6 OpenAI added a Meetings plugin to ChatGPT for Mac that turns calls into notes and action items. Google's differentiator is blunt — everyone else uploads your meeting audio to the cloud; Foresight doesn't.
For now Foresight is explicitly experimental, optimized for Apple Silicon Macs, and works with local files as well as Google Drive. Google is also using the same EmbeddingGemma 2 model to demo Instant Media Search and Video Moments Finder in its AI Edge Gallery app, and says ML Kit support for Android — with NPU acceleration where available — is coming in the coming weeks. TechCrunch's read: Foresight is mainly a showcase for what the Gemma family of offline models can do. But if users bite, a consumer-grade version folded into the Gemini app would be the obvious next move.