Google has put a working answer to the enterprise meeting-recording taboo on the Mac App alternative channel. AI Edge Foresight, a free experimental app released by the Google AI Edge team in early October alongside its EmbeddingGemma 2 model and widely reported this week, takes meeting notes entirely on the user's computer: it captures system audio and microphone input, transcribes the conversation, and turns shorthand bullets into polished notes — without joining the call as a bot and without sending a single byte to a server.
The interface follows the split-screen pattern that Granola popularized. Users jot quick notes on one side while the app generates fuller notes on the other, drawing on the live transcript. A Gemma 4-powered assistant is available for questions during and after the meeting, and a "Live Assistance" feature surfaces context cards with citations mid-call, pulling from past notes and documents.
The knowledge base is local-first as well. Foresight indexes folders on the Mac and Google Drive sources (indexing the latter still requires a connection), accepting PDFs, Google Docs, Microsoft Office files, plain text, Markdown and web bookmarks, with semantic search across all of it. When a meeting touches a topic covered in the library, the assistant can answer in real time. Google's pitch is that audio, transcripts, notes and files never leave the machine — the reason teams in banking, healthcare and government, where cloud transcription tools are routinely blocked, might actually install this one.
Under the hood is EmbeddingGemma 2, the 740-million-parameter open embedding model built on the Gemma 4 architecture and released under Apache 2.0. It carries an 8,000-token context window — enough for roughly 29 images or 5.5 minutes of audio — and Google says the text-only version needs about 191MB of RAM on a Pixel 11 Pro, with the full multimodal model around 567MB. The company reports a jump from 68.76 to 78.68 on the MTEB Code benchmark, a vendor claim pending independent verification. Google has not named the speech recognition model doing the transcription.
Practical constraints matter for anyone considering deployment. The app is roughly a 160MB direct download from Google rather than a Mac App Store listing, requires an Apple Silicon Mac — Intel Macs are excluded, and there is no Windows or iPhone version — and Google has not published minimum RAM or macOS version requirements. Recording consent is left to the user, which is a genuine compliance question in jurisdictions that require all parties to agree before a conversation is recorded.
The competitive context is crowded and getting more so. Granola built its business on the exact interaction Foresight now offers for free; Wispr and Calendly launched note-takers in recent months; and Superhuman — formerly Grammarly — acquired YC-backed note-taker Fathom just last month. What distinguishes Google's entry is what it owns downstream: the calendar, the documents, the video calls and the distribution. A Gemini-powered version integrated across meeting apps looks like the obvious next step.
Read as a product, Foresight is a modest experiment. Read as a signal, it is one of the clearest data points yet that the default architecture of AI productivity software is drifting from cloud API plus subscription toward on-device inference plus free acquisition. Sub-1B models that can run retrieval, embeddings and generation locally turn "your data never leaves the laptop" from a compliance concession into a product feature — and every SaaS note-taker charging per seat for cloud transcription now has a free, offline incumbent breathing down its neck.
Comments (0)
Log in to join the discussion
Log InNo comments yet