The open-source app reads focused-window text into local Markdown, giving AI agents a work history without storing a visual replay.
By RuntimeWire Staff · Published
Primary source: GitHub
Why it matters #
Ambient Context tests whether useful AI memory needs costly screen archives at all, or whether local, inspectable text can supply enough context for agents.
Cameron Smith (@cameronsmith) released Ambient Context on August 25th, turning the text visible in a Mac's focused window into a daily Markdown record that an AI coding agent can read later. The early open-source build avoids the screenshots, audio recordings and proprietary memory layer that have become standard ingredients in products promising to remember everything a user does. (github.com)
Smith is a product manager in Brisbane, Australia, whose GitHub profile lists him as head of ecommerce at Bloody Good Tests. His personal site says he spent more than 12 years building digital products and previously led product work at Replica Studios, an AI voice startup, where he helped design and launch B2B software. Ambient Context applies that product discipline to a narrower problem: giving an existing AI agent usable memory without asking the user to adopt another assistant, account or cloud service. (github.com)
The result is deliberately plain. Ambient Context checks the macOS Accessibility tree every few seconds, reads the text exposed by the active window and appends new material to one Markdown file per day. Each block includes a time range, application name and, when available, the document path or page URL. Repeated lines are removed so the file remains compact enough to pass to Claude Code or another agent as working context. (github.com)
A user could then ask what they worked on Tuesday, reconstruct the state of a project or draft a standup report. Ambient Context does not provide the model that answers those questions. It produces the record and leaves interpretation to software the user already runs.
Memory as a file, rather than a service
Smith's central product decision is the storage format. Ambient Context writes plaintext into a folder selected by the user. There is no account, server, telemetry or bundled model in the current build, according to the project's README, which says the software makes no network calls. The default folder sits outside ~/Documents
to reduce the chance that iCloud automatically syncs the record. (raw.githubusercontent.com)
That architecture gives users a memory layer they can inspect with a text editor, search with standard command-line tools, move between models or delete without navigating a vendor dashboard. Ambient Context also writes an AGENTS.md
file explaining the record's structure to whatever AI tool reads the folder.
The simplicity does not eliminate the privacy problem. Any continuous record of on-screen work can collect sensitive material, even when it omits pixels. Smith says Ambient Context excludes password managers and private-browsing windows, skips secure password fields and scrubs patterns resembling credentials, API keys and payment-card numbers before writing. It also ignores background windows, minimized windows, other displays and activity while the Mac is locked. Those protections are project claims attached to an unsigned 0.1.0 build, rather than guarantees established through a disclosed security audit. (github.com)
The data remains local plaintext, which makes ownership straightforward and device security consequential. A user, backup utility or other application with access to the selected folder can read the captured history.
The narrow capture model is the bet
Ambient Context is entering a computer-memory category already filled with products that collect much more. Screenpipe combines accessibility data with screen frames, OCR and audio transcripts, then exposes that history through a local API and MCP server. Littlebird says it reads on-screen text and adds its own assistant for searching work, meetings and projects; Littlebird announced an $11 million round led by Lotus Studio in March 2026. Microsoft's Recall stores searchable screen snapshots on supported Windows PCs. (littlebird.ai)
Smith has taken the opposite approach to feature coverage. Ambient Context watches one focused window, retains text and produces files for other tools. The trade offers lower storage requirements and avoids maintaining a permanent image or audio archive. It also means the product's usefulness depends on applications exposing meaningful text through Apple's Accessibility API.
Smith documents that weakness directly. Chromium and Electron applications can take several seconds to populate their accessibility trees, leaving early captures thin. GPU-rendered terminals including Kitty and Alacritty may expose little or no text, while Terminal.app and iTerm2 work better. Images, diagrams, video and canvas-rendered interfaces disappear from a text-only memory. The project's compatibility census asks testers to report which applications return complete, partial or empty results. (raw.githubusercontent.com)
That limitation defines the launch. Ambient Context is a test of whether enough professional work already exists as accessible text to make screenshots unnecessary for everyday recall. Developers, product managers and other workers who spend their days in browsers, documents, chat applications and conventional terminals are the clearest early users. Design, video and graphics-heavy workflows will leave larger holes.
An intentionally unfinished release
Ambient Context currently requires macOS 14 or later on Apple Silicon. It is unsigned and has no notarized download, so users must install Node, Rust and Apple's command-line tools, clone the repository and compile the application themselves. Smith says Apple Developer enrollment is in progress and that a signed release would add a single update check against GitHub. (github.com)
Those requirements keep Ambient Context firmly in tester territory. They also give Smith room to validate the underlying capture mechanism before building distribution, an assistant interface or a business around it. The repository does not establish Ambient Context as a venture-backed startup, and Smith has presented the work as a public software project rather than a funded company.
The first release still carries a coherent product thesis. AI agents are increasingly capable of acting on a user's work, but their context often begins with whatever someone remembers to paste into a prompt. Smith is betting that a chronological folder of ordinary Markdown can close much of that gap. Ambient Context will have to prove that text gathered through accessibility trees is complete enough to be useful and selective enough to be trusted. For now, it offers the memory category a refreshingly inspectable artifact: a folder full of files the user can actually open.