ChatGPT turns your Mac activity into an AI memory timeline

Remember the collective freakout when Microsoft unveiled Recall and started screenshotting everyone's desktops? OpenAI is now trying its own take on total system memory, but with a much lighter touch.

The company just launched Computer History for the ChatGPT macOS desktop app, rolling it out to Pro, Business, and Enterprise users. The feature quietly logs your activity across desktop apps and websites, transforming your daily grind into a searchable timeline and persistent memories that both ChatGPT and Codex can reference.

Clicks and keystrokes, zero screenshots

Instead of taking continuous visual snapshots like Windows Recall, Computer History captures an interaction-event stream. It watches clicks, typing, keyboard shortcuts, and app switches through macOS accessibility features, pulling textual context directly from the system.

OpenAI rebuilt the system from scratch to replace its earlier Chronicle preview. It explicitly does not capture screenshots, record webcam or system audio, or touch incognito browser tabs. As OpenAI product lead Ari Weinstein noted on X, private browsing sessions are completely ignored.

Locally, the event stream is cached inside a sandboxed macOS App Group for up to 48 hours before being wiped. Ephemeral Codex sessions process the raw events into human-readable Markdown files stored locally in your home directory. OpenAI states that server-side event processing does not retain data for model training.

In a quick demo, OpenAI's Dominik Kundel showed the app effortlessly identifying the last document he edited, confirming whether it had been shared with teammates in Slack, and drafting a standup recap of his morning.

Keeping the kill switch within reach

The feature is strictly opt-in and off by default. In enterprise workspaces, admins must first whitelist the role before any individual employee can toggle it on. Users must also have ChatGPT Memories enabled.

Granular controls let you whitelist or blacklist specific applications and URLs, which is vital if you handle sensitive financial sheets or private chats. A persistent macOS menu bar icon lets you pause collection instantly or wipe history across custom windows, from the last ten minutes to an entire day.

OpenAI notes that the summarizing process consumes standard tokens. It also flags a real security drawback: exposing system context heightens prompt injection risks if you land on a malicious webpage.

Availability reports conflict slightly. A third-party tracker indicated the feature was unavailable across the European Economic Area, Switzerland, and the UK, but OpenAI's official technical documentation explicitly lists those regions as supported.

Why it matters

We think OpenAI found a clever compromise in the battle for desktop context. By abandoning screenshots for accessibility event logs, it dodges the creepy surveillance backlash while still giving you an assistant that genuinely knows what you did before lunch. For heavy knowledge workers who waste hours piecing together standups or looking for lost tabs, it is definitely worth trying with a tight app exclusion list.

However, tracking keystrokes and accessibility trees comes dangerously close to keylogging, even with the best privacy safeguards in place. Relying on remote servers to summarize that stream requires immense trust, and the acknowledged prompt injection surface means letting an autonomous agent parse third-party web content from your active session remains a gamble.

For the broader industry, this sets a practical blueprint. It puts pressure on Apple and Microsoft to rethink heavy-handed screen capture models, proving that desktop AI agents can achieve deep cross-app workflows using lighter, text-first system telemetry.