OpenAI launches 'Computer History' for macOS, swapping screenshots for event tracking
The new opt-in feature records clicks and typing to build AI memories, though security critics warn of unencrypted local storage.
OpenAI has released "Computer History," an opt-in feature for the ChatGPT macOS desktop app designed to give the AI a searchable timeline of user activity. The tool is currently available to Pro, Business, and Enterprise users, though it remains unavailable in the UK, Switzerland, and the European Economic Area.
Rather than capturing visual data, the system utilizes macOS accessibility features to record specific interaction events. This includes tracking mouse clicks, typing, keyboard shortcuts, and switches between different applications. According to OpenAI, these events are stored locally for up to 48 hours before being sent to company servers to be processed into "memories." OpenAI states that these events are not retained after processing nor used for model training. To limit the scope of surveillance, the feature does not capture screenshots, screen recordings, microphone input, system audio, or activity within private-mode browsing.
A shift from visual capture
Computer History is a complete rebuild of "Chronicle," a previous research preview that relied on periodic screenshots to provide context to the AI. By shifting to an event-stream model, OpenAI aims to reduce token usage and improve privacy by eliminating the need for visual captures. This new iteration allows the AI to better understand the technical context of a user's workflow by tracking the sequence of interactions across the operating system.
Security and privacy concerns
While OpenAI frames the feature as a productivity tool for automating workflows and resuming interrupted tasks, the implementation has drawn sharp criticism. Critics have described the system as a keylogging and event capture mechanism that creates a detailed log of user behavior.
A primary security concern involves how the data is handled on the user's machine. OpenAI's own documentation acknowledges that local memory files—stored as plain-text Markdown—are not encrypted. This means any other program running with the same macOS user permissions may be able to access the sensitive activity logs. Furthermore, the reliance on an event stream increases the risk of prompt injection, where malicious content within a website or app could potentially manipulate the AI's memory by feeding it crafted interaction data.
The road ahead
As AI agents move toward more autonomous "long-term memory," the industry is watching how users balance productivity gains against these systemic vulnerabilities. The shift toward event-based tracking represents a broader trend in AI development, moving away from heavy visual processing toward lightweight, metadata-driven context. It remains to be seen if OpenAI will implement encryption for local storage or if the feature will eventually expand into the restricted European markets following further regulatory review.