OpenAI started rolling out Computer History in the ChatGPT macOS desktop app on 13 August, replacing the screenshot-based Chronicle preview with something more granular.
What it records
Rather than periodic screenshots, the feature builds an interaction-event stream — clicks, typing, keyboard shortcuts and app switches — read through macOS accessibility APIs, then summarises it into ChatGPT's memory. OpenAI is explicit about the exclusions: no screen images, no microphone, no system audio, no private-mode browsing.
The line in its own documentation
Raw events sit on the machine for up to 48 hours; the summaries derived from them last longer. OpenAI's documentation concedes the local files can contain sensitive information and are not encrypted by the feature — meaning any other program running as the same macOS user can read them. A plaintext record of everything typed is a well-understood target for infostealer malware.
The controls
It is opt-in and off by default, available to Pro, Business and Enterprise users on macOS only, with no API access. Availability in the EEA, Switzerland and the UK is deferred to a later phase — the standard sequencing for a feature whose legal basis under GDPR has not been settled.
The second-order risk
Everything the feature captures becomes long-term context for a model that also takes actions. Text on a visited page or inside an application can now enter memory and be read back later as instruction — a prompt-injection surface that widens exactly as agentic features ship.
Where this sits
Microsoft spent a year rebuilding Recall after a similar backlash, eventually encrypting its store and gating it behind Windows Hello. OpenAI has taken the opposite trade: less data captured, but stored in the clear.