OpenAI's Computer History Logs Your Activity as Structured Events, Not Screenshots — and Lets ChatGPT Use It as Context
OpenAI has introduced Computer History, an opt-in feature in the ChatGPT macOS desktop app that records what you do across applications and websites and turns it into a searchable timeline. ChatGPT and Codex (OpenAI's coding tool) can then reference that timeline when you ask a question. The feature appeared in the ChatGPT desktop app changelog on August 13, 2026, and is available to ChatGPT Pro, Business, and Enterprise customers on macOS.
The key design choice is what Computer History captures — and what it doesn't. Rather than taking periodic screenshots, it logs what OpenAI calls "events": discrete actions like clicks, keystrokes, and interactions within an application. No images, videos, or audio are recorded. The resulting timeline works as a form of contextual memory. When you issue a prompt, ChatGPT or Codex can draw on your recent activity to figure out what you mean, suggest ways to automate a task, or pick up something you left unfinished.
A demo video by Dominik Kundel, from OpenAI's Developer Experiences team, shows the workflow in action. In the recording, the app identifies the last document Kundel edited, checks whether it was shared with colleagues via Slack, and delivers a recap of his morning. The point is not to transcribe everything you did but to give the model a structured understanding of your activity across the whole operating system — not just inside the ChatGPT window.
Computer History is off by default. Pro users can turn it on themselves. Business and Enterprise administrators can manage it for their organizations. Individual users can exclude specific apps and websites from tracking and delete individual entries from the timeline. Ari Weinstein, Product and Engineering manager at OpenAI, stated on X that the feature will automatically ignore content in incognito or private browser tabs.
OpenAI's official X account described the goal plainly: with Computer History, "future interactions feel more personalized and require less explanation." The feature pulls context from selected apps and websites into ChatGPT and Codex, turning activity into memories and a timeline that persists across sessions.
The architectural difference from Microsoft's Windows Recall is worth understanding. Recall, which launched to significant controversy, stored periodic screenshots locally and let users search through them — raising concerns that sensitive information could be captured and later accessed by someone else on the machine. By logging structured events instead of pixel data, Computer History reduces the surface area for certain types of leakage. You cannot extract a screenshot of a banking session from an event log. The trade-off is that structured event data is harder to audit visually. A screenshot is immediately readable; a list of keystroke and click events requires tooling to interpret, which places more weight on the integrity of OpenAI's own processing and deletion controls.
The opt-in framing is consistent with how OpenAI has handled its most privacy-sensitive features. The granular exclusions — per-app, per-site, per-entry — give technically literate users the controls they would expect. Admin-level management for Business and Enterprise tiers aligns with what IT teams responsible for fleet-wide deployments would need before enabling something that broadly captures workstation activity.
The broader context here is that Computer History sits at the intersection of two trends that have been converging for the past two years. Context windows — the amount of text a model can hold in memory during a conversation — have expanded into effectively unbounded territory. At the same time, there is a push toward agent-like workflows, where models act on a user's behalf across multiple tools. A model that can remember what you did this morning is a prerequisite for one that can take over where you left off. Whether that capability is net positive depends entirely on execution and trust. The technical approach — structured events, no screenshots, incognito awareness, granular controls — is a reasonable design for the problem. The remaining question, as with any feature that logs user behavior at this granularity, is what happens to that data downstream: how long it persists, whether it feeds model training, and what guarantees users have that deletion is irreversible. OpenAI says user actions are turned into training data. The controls are present. The burden of living up to them is ongoing.
