OpenAI has introduced Computer History, an optional feature in the ChatGPT desktop application for macOS that allows ChatGPT and Codex to reference selected activity across apps and websites. The company says the feature is designed to help users continue tasks without repeatedly explaining what they have already done.
How the feature works
Computer History records interaction events exposed through macOS accessibility features, including clicks, text input, keyboard shortcuts and switches between applications. Unlike OpenAI’s earlier Chronicle research preview, the feature does not rely on screenshots or screen recordings. It also excludes microphone input, system audio and private-browsing activity, according to OpenAI.
The feature converts selected activity into a searchable timeline and plain-text Markdown memory files. These memories can later provide context to ChatGPT and Codex during conversations or tasks, potentially allowing the systems to recall the research, documents or applications a user worked with earlier.
Computer History is switched off by default. It is currently available to ChatGPT Pro, Business and Enterprise subscribers, while Business and Enterprise users require permission from their workspace administrator before opting in. OpenAI said the feature is not yet available in the European Economic Area, the United Kingdom or Switzerland.
User controls and privacy
Users can decide which applications and websites are included, pause collection from the Mac menu bar and inspect or delete timeline entries. The controls also allow users to clear activity from the previous 10 minutes, hour, day or the entire history.
OpenAI says interaction events are temporarily stored on the Mac for up to 48 hours. They are then processed on the company’s servers to generate memories, after which OpenAI says the events are not retained or used to train its models. The resulting memory files remain on the user’s Mac until they are deleted.
The feature’s privacy model nevertheless creates new risks. Activity may reveal sensitive work patterns, research topics, communications or information entered into applications. OpenAI advises users to exclude apps containing confidential data and pause collection when communicating with people who have not provided express consent.
Prompt-injection concerns
OpenAI has also warned that Computer History can increase exposure to prompt-injection attacks. Malicious instructions embedded in a webpage, document or application could potentially enter the context later used by ChatGPT or Codex.
This is particularly significant as AI systems move from answering questions to carrying out tasks. Users should treat remembered activity as potentially untrusted context and review suggested actions before allowing an agent to modify files, send messages, access accounts or execute workflows.
From memory to automation
Computer History could eventually become more than a recall tool. OpenAI says repeated activities may be used to propose reusable skills or scheduled automations that Codex can build from a user’s recorded workflow.
For example, a journalist who regularly gathers information from several websites, checks facts in a spreadsheet and prepares a briefing document could use the feature to help Codex understand that sequence. In principle, the agent could later assist with parts of the process without requiring the user to describe every step again.
The convenience comes with an important trade-off: the more an agent understands about a person’s working habits, the more carefully its permissions and outputs must be managed. For professional users handling unpublished material, personal information or confidential sources, a cautious rollout—using narrow app permissions, regular history reviews and manual approval for external actions—will be essential.
