The Pulse
Google Announces Offline Foresight Meeting Companion for Mac
Google’s experimental Foresight app uses local AI models to assist with meeting notes and search across transcripts and personal files on Mac, including when offline. The announcement also details two mobile search demos and tools for devel
AI.info Team · Written with AI ·
Google announced Google AI Edge Foresight for Mac on Oct. 6, 2026, describing it as an experimental meeting companion that processes information locally. The app integrates with a Mac’s system audio and microphone, which Google says lets it work with any meeting platform, even without an internet connection. The post says Foresight uses EmbeddingGemma 2 and Gemma 4.
Foresight brings local meeting assistance to Mac
Google says Foresight can enrich shorthand notes with details retrieved from the current conversation, help users find images, documents, transcripts and notes using natural language, and assist with answering questions during meetings or drafting longer notes. Its searches run across a personal knowledge library, with text, audio and visual data mapped into a shared vector space. Google says the information stays on the device, the app can continue working without a network connection, and it has no cloud subscription costs.
The post introduces the app this way:
“We are excited to announce the launch of Google AI Edge Foresight for Mac, an experimental app that brings a context-aware meeting companion directly to your device.”
Google AI Edge Team and Android ML Team, the post’s authors
EmbeddingGemma 2 supports multimodal search
Google describes EmbeddingGemma 2 as an open-weight, multimodal embedding model with 740 million parameters. It maps text, images, video frames and audio into a single vector space. The post says that combining those modalities can reduce the latency and memory overhead involved in chaining separate image-captioning, speech-to-text and text-embedding models. It does not specify which tasks in Foresight are assigned to EmbeddingGemma 2 and which to Gemma 4.
Embedding models represent information as vectors that can be compared for similarity. In Foresight, Google says, this approach supports searches across different kinds of personal files as well as meeting transcripts. The post presents note enhancement and cross-modal retrieval as app features, without detailing a division of responsibilities between the two models.
Two new mobile search demonstrations
The announcement also adds Instant Media Search and Video Moments Finder to Google AI Edge Gallery, a showcase app for trying on-device AI models. Instant Media Search finds local images and videos using natural-language queries or example images. Google says it stores media embeddings in a local SQLite database and ranks results by cosine similarity, updating results as a user types.
Video Moments Finder searches local video recordings for visual moments without transcribing audio or generating intermediate text captions. After a user selects a video, the on-device engine indexes visual keyframes alongside audio chunks. A query such as “dog catching a frisbee” can then be matched against video frames to identify relevant timestamps.
Tools for developers building local search
Google says EmbeddingGemma 2 will be available through ML Kit on Android in the coming weeks, with NPU acceleration intended to improve performance on supported devices. For cross-platform development, the post describes MediaPipe Tasks support for the model’s Embedder and Semantic Retriever components. The tools are intended to handle steps such as image resizing, normalization and multimodal tokenization, and can be used for apps targeting iOS, macOS, Windows, Linux or the web.
The post also describes deploying the model with LiteRT, which supports execution across CPUs, GPUs and NPUs. Google reports that visual embeddings took as little as 37.3 milliseconds per image on a MacBook M5 Pro GPU in its stated test. The company says developers can try the mobile demonstrations in Google AI Edge Gallery and download Foresight for Mac.