OpenAI’s ChatGPT Now Records Every Keystroke on Your Mac

Apple introduced the Macintosh in 1984 with a promise that computers should work the way people do. Forty-two years later, OpenAI has taken that idea to an extreme few anticipated. Its new Computer History feature for ChatGPT logs clicks, keystrokes, keyboard shortcuts and app switches on macOS devices. The system turns those inputs into searchable memory the model can reference in future conversations.

The launch landed quietly on August 16, 2026. The Next Web broke the details first. OpenAI built the capability through macOS accessibility frameworks, the same pathway screen readers and automation tools have long used. No screenshots. No video. No audio. Just the raw events of human interaction with the machine.

But the implementation carries trade-offs that industry observers immediately flagged. The resulting memory files sit on the local drive as unencrypted plain text Markdown. Any application running under the same user account can read them. OpenAI acknowledges the risk in its own documentation.

This marks a sharp pivot from an earlier experiment. The company previously tested Chronicle inside its Codex coding assistant. That version captured screen images to give the model visual context. Users pushed back. Privacy concerns mounted. Computer History replaces those images with input events. The change reduces visual data collection yet still creates a detailed record of user behavior.

And the timing fits a larger pattern. OpenAI rolled out its Computer-Using Agent model in January 2025, powering the Operator agent that could navigate browsers and interfaces on a user’s behalf. By July 2025 the company integrated those abilities into ChatGPT Agent, which reached 10 million users within weeks according to Yahoo Finance. The desktop Computer Use tool announced in July 2026 lets the model act directly on a user’s machine, clicking, typing and moving files in the background. Computer History supplies the memory layer that makes those actions more informed.

OpenAI’s own blog post on July 9, 2026 described the vision. “On desktop, Computer Use lets ChatGPT use your computer on your behalf to execute tasks in the background across your apps, tools, and browser—clicking, typing, and moving files where they need to go,” the company wrote in its announcement. ChatGPT Work, the broader agent product, can stay with a project for hours, breaking complex goals into steps and completing them independently.

Computer History feeds directly into that ambition. A user might ask what document they edited before lunch. The model pulls the exact sequence of keystrokes, app switches and file interactions. It can surface a lost spreadsheet or recap Slack threads reviewed that morning. Early demos shared on X showed the assistant locating recent work, checking messages and offering context without the user providing explicit prompts.

Yet the feature arrives with deliberate restrictions. It is unavailable in the European Economic Area, Switzerland and the United Kingdom. OpenAI cited regulatory reasons. For enterprise and business customers, administrators must enable it first. Individual users then grant consent. The Memories toggle must be activated separately. Users can maintain lists of included and excluded applications and websites. A menu bar icon allows pausing at any moment.

Retention periods stay short. Temporary event logs remain on the device for 48 hours before automatic deletion. OpenAI processes them on its servers but claims it does not retain the raw data afterward except where legally required. The memory files themselves do not train future models. Still, when those memories appear in chat conversations, the chat content may become training data depending on user settings.

The company warns of limitations. Prompt injection risks rise because the model receives detailed context from user actions. Communication apps pose special problems. Recording conversations without all parties’ consent creates legal and ethical exposure. OpenAI advises against using the feature with health, financial or personal data applications. These cautions echo long-standing industry struggles with always-on recording tools.

Privacy advocates drew immediate parallels to Microsoft’s Recall feature, introduced in 2024 for Copilot+ PCs. Recall took periodic screenshots and stored them locally. Security researchers demonstrated how easily the database could be accessed. Microsoft delayed the rollout and added encryption after the backlash. Computer History avoids screenshots but stores event logs in plain text. The similarity in risk profile is hard to ignore.

But here’s the difference. Recall aimed at search and recall through visuals. Computer History focuses on building procedural memory. It watches patterns. It notices repeated sequences of actions. The system can suggest turning those patterns into automated skills or scheduled tasks. Codex, now integrated into the main ChatGPT desktop app, can generate code to replicate the observed workflow. The loop from observation to automation closes faster than before.

Analysts see this as the logical next step for agentic systems. The Verge reported in July 2025 that ChatGPT Agent operates on its own virtual computer to complete multi-step work. Giving that agent a persistent record of the user’s actual behavior makes its suggestions more relevant. It reduces the need for users to describe their processes in detail every time.

Adoption numbers tell part of the story. OpenAI reported nearly 100 percent of its internal teams use the agent products for complex projects. Customer examples include Zapier automating lead reviews, RingCentral handling product support tickets, Virgin Atlantic analyzing flight data and NVIDIA preparing for major events. The addition of desktop Computer Use and now Computer History expands the surface area where these agents can act with meaningful context.

Security considerations extend beyond local storage. Enterprise deployments include admin controls over which tools the agent can access, what actions it may take and how much it can spend. A Compliance API provides visibility. Automatic review mechanisms flag sensitive operations. These safeguards matter because the agent now sees the same interfaces employees use to access corporate systems.

Critics argue the technology still falls short of seamless execution. Tests of earlier computer-use agents showed they excel at simple browser tasks but stumble on complex, stateful workflows. Error rates remain high enough that human supervision stays necessary. Computer History may improve performance by providing better context, yet it does not solve the underlying challenges of visual understanding and reliable action selection.

The plain text storage choice raises eyebrows among security professionals. Local files in Markdown format offer easy debugging and searchability. They also create an obvious target. A compromised account grants an attacker the complete history of user interactions across permitted applications. OpenAI’s documentation flags this exposure. The company appears to have prioritized transparency and functionality over hardened encryption in the initial release.

So what comes next? Industry watchers expect tighter integration between the memory system and the agent capabilities. Scheduled tasks could run with awareness of past user preferences derived from keystroke patterns. Automation suggestions might grow more sophisticated as the model accumulates months of interaction data. The boundary between observation and action continues to blur.

OpenAI faces the same tension that has defined personal computing since the Macintosh debuted. Tools that understand users deeply also watch them closely. Features that increase productivity can diminish privacy. The company has responded with layered consent, granular controls and geographic restrictions. Whether those measures satisfy regulators, enterprises and individual users will determine how widely Computer History spreads.

For now, the feature exists as an opt-in experiment available to ChatGPT users outside Europe on supported desktop platforms. It represents one more step toward AI systems that don’t just answer questions but remember how their human partners operate. The keyboard has always been the primary interface between person and machine. OpenAI just gave the machine a better way to read what fingers have written there.