Google's Foresight runs offline AI meeting notes on macOS

Google released a free, experimental macOS application on October 8, 2026 that transcribes and summarizes meetings without transmitting audio to remote servers, according to TechCrunch. The app, called Google AI Edge Foresight, runs entirely on-device using the company's EmbeddingGemma 2 model — a compact language model optimized for Apple silicon (Apple's custom ARM-based processors found in M-series chips).

The release positions Google against cloud-reliant note-taking services Granola and Wispr Flow, both of which require internet connectivity to perform speech-to-text conversion and large language model summarization. Foresight executes the entire pipeline — audio capture, transcription, summarization, and note generation — on the user's Mac, a configuration that eliminates recurring cloud-compute costs and removes audio data from third-party servers.

What does Foresight actually do?

The application combines meeting transcription with a manual note-taking surface, in the same vein as Granola and Wispr Flow. Users open Foresight during a call, and the app records system audio while displaying an editable notes panel. Google says the tool will transform shorthand bullet points typed during a meeting into what the company calls "polished notes," structured prose built from the transcript.

Two workflow modes distinguish Foresight from a passive recorder:

• Live transcription. The app converts meeting audio to text in real time, displayed as a scrolling caption log.
• Post-meeting transformation. Users type bullet fragments during the call, and Foresight rewrites them into full sentences after the meeting ends, anchored to the transcript for factual grounding.

The "polished notes" framing reflects a design choice to keep humans in the editing loop rather than generating autonomous summaries. The transcript remains visible, allowing the user to verify any rewritten point against the source recording.

How does the on-device architecture work?

Foresight runs EmbeddingGemma 2, a model Google published for local inference, on macOS hardware. The "Embedding" prefix refers to a vector encoding process — the transformation of text into numerical representations that capture semantic meaning. The "Gemma" suffix denotes Google's open-weight model family. The "2" marks the second generation.

Key technical points drawn from Google's developer documentation:

• Local inference. All model execution occurs on the user's CPU and GPU, not on Google's servers.
• EmbeddingGemma 2 design. The model produces vector representations of text — numerical fingerprints a retrieval system can search — suitable for retrieving relevant transcript segments without cloud lookup.
• macOS exclusivity. The initial release targets macOS only. Windows and Linux versions are not listed at launch.
• Free of charge. Google has not announced pricing tiers or subscription plans. Users can download the app without payment.

The combination of local vector embeddings (numerical text fingerprints used for similarity search) and local generation enables Foresight to answer questions about past meetings without round-tripping data to Google's infrastructure. A user can ask "what did we decide about the launch date?" and the app searches the locally stored transcript embeddings to surface the relevant passage.

Why does offline processing matter?

The shift to on-device AI addresses three structural concerns that have shaped enterprise software purchasing decisions since 2023:

1. Data residency. Audio and transcript content never leave the user's machine, simplifying compliance with internal data-handling policies and regulated-industry requirements such as HIPAA (the US Health Insurance Portability and Accountability Act) and GDPR (the EU General Data Protection Regulation).
2. Operating cost. Cloud transcription typically bills per minute of audio. Local processing carries only the one-time hardware cost of the user's machine.
3. Latency. On-device inference avoids the round-trip to a remote data center, producing transcription output within the same network environment as the user's other applications.

For corporate IT teams, the privacy dimension carries the most weight. A transcribed meeting may contain customer names, pricing discussions, or pre-release product details — categories of information that legal departments routinely flag for restricted transmission. Foresight's offline architecture removes that transmission step entirely.

How does Foresight compare to existing note-takers?

Granola and Wispr Flow have established a market category sometimes labeled "AI meeting notetakers." Both operate on a similar architecture: capture audio, transcribe via cloud API (a remote service that converts speech to text using server-side models), and summarize with a hosted large language model.

The defining differentiator is offline capability. Granola and Wispr Flow require connectivity for both speech recognition and language modeling. Foresight performs both locally, removing a dependency that has caused both competitors reliability complaints when networks fail.

What are the limitations?

The source material on Foresight is limited. The app arrived as an experiment rather than a finished product. Several constraints apply:

• macOS only. Windows and Linux users have no equivalent at this stage.
• Apple silicon required. EmbeddingGemma 2 inference depends on the neural engine (a dedicated hardware block for matrix math) present in M-series chips. Intel-based Macs may not deliver acceptable performance.
• Experimental status. Google has labeled the release experimental, signaling that features may change and support may be withdrawn.
• Unknown accuracy ceiling. The published material does not report word error rate (the standard speech-recognition accuracy metric, expressed as a percentage of misrecognized words) for Foresight's transcription output.

Where does Foresight fit in Google's product strategy?

The app belongs to Google AI Edge, a developer initiative Google has used to distribute models that run outside Google's data centers. EmbeddingGemma 2's presence in the Foresight stack aligns with Google's broader push toward "local-first" AI — the principle that core inference should occur on the user's hardware whenever feasible.

The release arrives in a competitive environment shaped by two pressures: regulators tightening data-localization requirements and enterprises demanding predictable AI costs. Foresight's offline architecture addresses both, though its experimental status suggests Google has not committed the engineering investment required for a fully supported commercial product.

What should readers watch next?

Three signals will determine whether Foresight graduates from experiment to production:

• Platform expansion. A Windows or Linux release would expand the user base. Absent that, Foresight remains a Mac-only release.
• Model evolution. EmbeddingGemma 3, if released, would likely reach Foresight first, raising accuracy ceilings.
• Pricing transition. Google has a documented pattern of releasing experimental features for free and introducing paid tiers upon graduation. Foresight may follow the same path.

The October 8, 2026 launch establishes a technical baseline: a free, on-device, macOS-native meeting notetaker backed by a public Google model. Whether Google treats the app as a strategic product or an academic demonstration will become clear when the first update ships.