Google's new meeting assistant makes a more specific promise than the usual "local-first" label: its AI inference, vector index, meeting audio and transcripts all stay on the Mac. The claim covers the whole pipeline. The experimental Google AI Edge Foresight app packages it into a downloadable meeting tool, giving developers a public example of which pieces of a retrieval-augmented generation system can now fit on one laptop, even as Google leaves implementation details unexplained.
Released for macOS this week, Foresight listens to system audio and the microphone, transcribes calls or face-to-face conversations, expands a user's shorthand and answers questions against imported work files. TechCrunch reported that the app resembles Granola in its side-by-side manual and generated notes. The sharper distinction is below the interface: Google says the meeting can continue, be searched and be queried with no internet connection.
A local claim that reaches past inference
Many AI apps use "local" to describe one stage while sending another stage to an API. Google's Foresight FAQ says all inference runs on the device and that personal knowledge sources, recordings, transcripts and notes are never transmitted to external cloud servers. It also says the vector embeddings engine runs locally. That last part closes an easy-to-miss route by which document contents or derived representations could otherwise leave the computer.
Foresight can ingest PDFs, Google Docs, Microsoft Office files, plain text, Markdown and web bookmarks. Users can point it at project folders, reference material, diagrams and a calendar, then retrieve that context through chat or automatic answer cards during a meeting. Google's page says files stay on the Mac, although it does not explain the network and authentication path used when a user connects Google Drive or adds a web bookmark. Offline inference and offline source acquisition are separate questions.
The app works across meeting services because it captures the Mac's own audio rather than joining as a bot. That also gives it substantial operating-system access. Apple places microphone and system-audio recording behind separate controls in Privacy & Security, and shows an orange indicator for microphone use or a purple one for system-audio recording. Apple's documentation lets users revoke those permissions per app. Local processing removes a cloud data path, but the laptop becomes the place where recordings, transcripts and the searchable index need protection.
Two models, with different jobs
The 740-million-parameter model in Google's announcement is not the component writing finished meeting notes by itself. EmbeddingGemma 2 converts text, code, images, video and audio into points in one 768-dimensional space. Foresight uses those representations to find relevant passages or media. A Gemma 4 model then supplies the generative assistant that expands notes and produces answers. Keeping retrieval and generation distinct is important when assessing both the app's behavior and its resource needs.
EmbeddingGemma 2 is modular. Its text path has 270 million parameters, while the optional vision and audio encoders add 170 million and 300 million respectively. Google reports about 191MB of active RAM for text-only weights and about 567MB for the full multimodal model on a Pixel 11 Pro in its developer announcement. Those figures describe the embedding component on Google's test phone, not Foresight's total Mac footprint. Google has not named the Gemma 4 variant inside Foresight or published the app's total memory, storage and energy use.
The shared embedding space cuts out some intermediate work. A developer can compare a text query directly with an audio segment or video frame rather than first generating a caption for every item. In Google's mobile gallery demo, media vectors are stored in a local SQLite database and ranked by cosine similarity as the user types. Foresight applies the same basic idea to meeting transcripts and work files, then grounds a separate generator with the retrieved material.
There are firm input limits beneath that design. The model card gives EmbeddingGemma 2 an 8,192-token context window and charges audio at 25 tokens per second, for a maximum of roughly 327 seconds when audio uses the whole window. An hour-long meeting cannot be embedded as one raw-audio input under those defaults. The app must index smaller units or use another processing stage, but Google has not published Foresight's chunking, transcription or retrieval pipeline. That missing detail will affect whether a question retrieves the right five minutes from a long call.
The useful parts are available to developers
Foresight is an experimental download optimized for Apple Silicon, yet the underlying embedding model has a broader path. Google released EmbeddingGemma 2 under the Apache 2.0 license and provides weights for Transformers and Sentence Transformers. Developers can load only the encoders they need: text and audio total 570 million parameters, for example, while a text-only search service can leave both media encoders out.
The output vectors can also be shortened from 768 dimensions to 512, 256 or 128. Google's model card reports a threefold storage reduction at 256 dimensions with a smaller benchmark drop than the 128-dimensional setting. There is a trap in the optimization: truncated vectors must be normalized again before cosine comparison. If they are not, the search still returns plausible scores while its ranking quality degrades. That is the sort of quiet failure a polished demo does not reveal.
Precision has another sharp edge. Google tells developers to run the model in bfloat16 or float32 and explicitly warns against float16, where the activation range can produce NaNs or silently damaged embeddings. Consumer AI hardware discussions often treat lower precision as a simple memory trade. Here, choosing the wrong 16-bit format can corrupt the retrieval layer without a clean crash.
Google is building higher-level routes around those weights. Its announcement says ML Kit support for Android is due in the coming weeks, with automatic model updates and NPU acceleration where available. MediaPipe Tasks is adding an embedder that handles resizing, normalization and tokenization, plus an on-device approximate-nearest-neighbor retriever. LiteRT remains the lower-level option for teams that want direct control across CPUs, GPUs and NPUs. These are announced integration paths, not evidence that Foresight itself is ready for managed workplace deployment.
Privacy does not answer the accuracy question
Google's privacy claim is unusually plain, but its public material leaves the quality of the complete app largely unmeasured. The EmbeddingGemma 2 model card publishes results across text, code, image, video and audio retrieval benchmarks. The Foresight page does not publish transcription error rates, retrieval recall, note fidelity, speaker separation results or end-to-end latency. It says the app has a light footprint and low battery drain without giving a test machine, meeting length or power measurement.
Those omissions matter during a live call. A wrong transcript can poison the index before retrieval starts, and a relevant passage can still be missed when a long meeting is divided into chunks. Even a correct retrieved passage can be summarized badly by the generative model. Foresight's product page says its proactive context cards include exact citations, which gives a user something to inspect, but Google has not published an evaluation of how often those citations support the answer shown.
The local design also moves administrative work rather than removing it. Teams still need rules for recording consent, retention, deletion and backup. The public FAQ reviewed for this article does not describe enterprise policy controls, encrypted export, automatic deletion or an audit log. Apple's permission switches can stop future microphone or system-audio access; they do not define what happens to meeting data the app has already stored.
Foresight enters a crowded category. TechCrunch notes that Wispr and Calendly recently introduced meeting note tools, while Superhuman acquired Fathom. Google may have an advantage in owning both the models and deployment libraries, but this release is labeled experimental and restricted to Apple Silicon. The download page does not give minimum RAM, a macOS version floor or reproducible performance targets, so buyers cannot yet compare its local cost with a cloud subscription on equal terms.
Foresight is most useful today as a concrete test of Google's edge stack. Watch for a named Gemma 4 configuration, measured transcription and retrieval results, documented storage controls, and the promised ML Kit and MediaPipe support. Those disclosures will determine whether developers can judge the system instead of the demo. Until then, the strongest fact is also the narrowest one: Google has put a meeting recorder, multimodal index and generative assistant on one Mac without requiring the meeting contents to visit its cloud.