Ambient Context Releases a Mac Work Record for External AI Agents

Ambient Context has reached its first public release with a deliberately narrow answer to AI memory: record the text in the Mac window a person is actively using, save it as local Markdown, and let an external agent interpret it later. The trade is clear. The app avoids screenshots and bundled model access, but its usefulness depends on what each application exposes through macOS accessibility tools.

A written trail from the window in front

Ambient Context is a macOS menu bar app that polls focused-window text through the Accessibility API every few seconds. It stores a time range, app name, window title, and, where available, a document path or page URL. Repeated lines are removed across the day to keep the resulting record compact.

What each day can produce

• Context: a written record of active-window activity, organized in daily files.
• Knowledge: six cited files covering People, Commitments, Threads, Products, Issues, and Reading.
• Notes: a generated daily write-up based on the timeline and cited knowledge base rather than raw text alone.

The model stays outside the app

The capture record is not answered by an assistant inside Ambient Context. Users connect a command-line agent, such as Claude Code, Codex, or opencode, to run the knowledge-base and note-writing workflows. The app itself holds no agent credentials and makes no model calls; an agent runs locally under the user’s own subscription or configuration.

Local storage narrows collection, not the privacy stakes

The app does not take screenshots, record video, run OCR, read background windows, or capture activity while the screen is locked. Its daily output, including the context record, cited knowledge base, and notes, is plain Markdown in a folder the user selects. That keeps the history inspectable and movable while preserving a written record of active work.

The project says it excludes password managers and private-browsing windows, skips secure password fields, and scrubs patterns resembling credentials, API keys, and payment-card numbers. Those controls reduce some obvious exposure routes, but the product is still designed to preserve a continuous text history of active work.

Coverage gaps set the practical boundary

The approach can only preserve text an application makes available to the Accessibility API. The project documents incomplete capture for some Chromium and Electron apps, GPU-rendered terminals including Kitty and Alacritty, and interfaces built from images, diagrams, video, or canvas rendering. A work history assembled this way may be easy to inspect, but it will not be a complete replay of a person’s day.

The public release requires macOS 14 or later on Apple Silicon and is MIT-licensed. The immediate test is whether a selective, text-only record gives outside agents enough dependable context to be useful without becoming another opaque archive.