OpenAI has rolled out Computer History, a new opt-in feature for the ChatGPT desktop app on macOS that records activity across permitted apps and websites, turning it into a searchable timeline and local memories for ChatGPT and Codex. The feature requires ChatGPT Memories to be enabled.
Users can ask questions such as what they were working on before a break, where they last saw a document or what they accomplished during the day. ChatGPT can use the activity timeline to identify the relevant file, conversation or website and then access that source directly.
OpenAI says Computer History replaces its earlier Chronicle research preview, but the two work differently. Chronicle relied on screenshots, while the new system records interaction events such as clicks, typing, keyboard shortcuts and app switching through macOS accessibility features.
“Computer History doesn’t rely on screen or audio capture. Instead, it captures interaction events like clicking, typing, app switches, and more, so it can capture memories faster and more efficiently,” OpenAI’s Dominik Kundel said in an announcement video.
From memory to automation
The more interesting part of Computer History may be what happens after ChatGPT learns a user’s routine. When the system spots a repeated workflow, it can suggest turning it into a skill or automation. Users can then ask Codex to build a reusable workflow based on the recorded activity.
That pushes ChatGPT beyond simply remembering previous work. It gives the system a way to learn how a person performs tasks and potentially reproduce those steps later. For businesses, that could make repetitive processes easier to automate, extending ChatGPT’s growing role across connected office apps, files and websites. For individual users, it could reduce the friction of returning to unfinished work or finding something they vaguely remember seeing.
The privacy trade-off
OpenAI has built several controls around the feature. Users choose which apps and websites contribute, can pause collection from the macOS menu bar and can delete individual entries or larger periods of history. Private-mode browsing is excluded automatically.
The company also says Computer History does not capture screenshots, screen recordings, microphone input or system audio. Temporary event files remain on the Mac for up to 48 hours before deletion after processing, while generated memory files remain locally until the user removes them.
But the local storage comes with an important caveat: OpenAI says the memory files are plain-text and are not encrypted by Computer History. Other programs running under the same macOS user account may potentially access them. OpenAI also warns that the feature increases the risk of prompt injection because malicious instructions on websites or inside applications could enter the context available to ChatGPT or Codex.
A new bet on personal AI
Computer History signals a broader shift in how OpenAI sees AI assistants. Instead of waiting for users to explain what they need, ChatGPT can build context from the work already happening around it.
That could make AI agents considerably more useful, but it also raises the cost of a mistake. The more an assistant knows about someone’s daily workflow, the more valuable its memory becomes, and the more carefully that memory needs to be protected.
For users, the practical choice is therefore less about whether Computer History is useful and more about where its usefulness outweighs the privacy and security trade-offs. OpenAI has made the feature optional and controllable, but its own warnings suggest that selective use may be wiser than simply switching it on everywhere.
The feature is available to ChatGPT Pro, Business and Enterprise users, although Business and Enterprise administrators must first grant access before employees can enable it. It is off by default and currently unavailable in the European Economic Area, Switzerland and the United Kingdom.
Read more: Computer History’s prompt-injection warning reflects a broader agent-security problem, as OpenAI’s GPT-5.6 tests found higher attack success rates when malicious instructions came through tools and external content.
