Core Event

OpenAI is adding a new feature called Computer History to the macOS ChatGPT desktop app: it records user actions like clicks and keystrokes on the computer and compiles them into a timeline that ChatGPT and Codex can query.

This means that when users ask ChatGPT a question, the system won’t rely solely on the current conversation — it may also reference the user’s prior activity trail on the computer. For example, it could help recover a recently edited document, determine whether a piece of work was already shared via Slack, or summarize what the user did during the morning. According to The Verge, this capability resembles Microsoft’s controversial Windows Recall, but OpenAI emphasizes that Computer History does not capture screenshots, nor does it collect images, videos, or audio — instead, it describes user activity through “events.” These “events” can be understood as operational records within applications, such as traces left by opening, editing, clicking, or typing.

How It Works: From Local Behavior to a Queryable Timeline

Computer History aims not simply to log a raw chronology, but to help ChatGPT learn how users work and provide more context-aware assistance based on that understanding. According to the original report, it could suggest automation flows and even pick up tasks the user left mid-way. Codex can also reference this timeline; Codex is OpenAI’s tooling or model system designed for programming tasks, commonly used for understanding code, generating code, or assisting developers with engineering workflows.

Currently disclosed information shows that the key capabilities of Computer History include:

• Recording event traces such as clicks and keystrokes to form a user activity timeline;
• Making the timeline available for ChatGPT and Codex to reference when the user initiates a request, supplementing context;
• Helping回溯 recent tasks, such as finding the last-edited document;
• Assisting in assessing workflow status, such as checking whether content has been shared via Slack;
• Generating activity overviews, such as summarizing what the user did in the morning.

Dominik Kundel, a member of OpenAI’s developer experience team, demonstrated these scenarios in a short walkthrough: the app located his most recently edited document, checked whether it had already been sent to someone via Slack, and provided a morning work recap. This demo illustrates that Computer History is more like adding a layer of “working memory” to an AI assistant, so it doesn’t have to wait for users to describe their background piece by piece.

Privacy Controls: Opt-In by Default, Not Opt-Out

The most sensitive aspect of a feature like this is clearly privacy. Click, keystroke, and app usage trails can reveal a great deal about work habits, project progress, and even personal information. Ari Weinstein, OpenAI’s product and engineering manager, stated on X that Computer History is an opt-in feature — not something enabled by default that users then have to disable. Users can also exclude specific apps and websites and delete individual entries when needed, giving them finer-grained control.

Additionally, Computer History automatically ignores content from incognito or private browser tabs. This is significant, since private browsing mode is typically understood by users as an environment that should not be recorded long-term. Nevertheless, “not taking screenshots” does not mean “no privacy risk.” Event logging, while more restrained than continuous screen capture, can still expose which tools a user has accessed, what tasks were handled at what times, and how different workflows interconnect.

Similarities and Differences with Windows Recall

The Verge drew a comparison between Computer History and Windows Recall for an obvious reason: both attempt to let AI answer questions and recover context through the user’s past computer activity. Windows Recall sparked controversy largely because it relied on a large volume of screenshots to build a searchable memory. OpenAI, in contrast, emphasizes that it does not collect images, videos, or audio — only events.

This design difference will shape user perception. Screenshots naturally contain more visible information, potentially capturing chat windows, document content, and web pages all at once; event logging is comparatively structured and theoretically easier to filter, delete, and govern with permissions. However, for average users, the real question remains transparency: exactly which events are recorded, how long they are retained, where they are processed, and how they are called upon by ChatGPT or Codex. The original article does not provide these details, so at present we can only confirm that OpenAI has disclosed mechanisms for opt-in enrollment, app and website exclusions, entry deletion, ignoring private browsing tabs, and not collecting images, video, or audio.

Industry Perspective: AI Assistants Are Evolving From “Chat Boxes” to “Working Memory”

Computer History reflects a clear direction for desktop AI assistants: moving from passively answering questions toward understanding the user’s work context, proactively suggesting automation, and helping bridge unfinished tasks. For everyday tech users, this could deliver tangible productivity gains — especially in work environments where multiple apps, documents, and communication channels run in parallel.

But the pace of adoption for this kind of capability ultimately depends on trust. For AI to become a true desktop assistant, it must prove it can strike a balance between “useful” and “restrained”: remembering what users need it to remember, while clearly steering clear of areas users do not want recorded. OpenAI’s choice to use events rather than screenshots and to make the feature opt-in is a step toward reducing controversy; what users need to see next are clear records of what’s captured, auditable history entries, and stable, understandable controls. The future competition in desktop AI may not just be about which model is smarter — it may also be about who manages memory, boundaries, and trust best.