IA · 16 August 2026 · 5 min read
OpenAI Introduces Computer History: ChatGPT Learns from Mac Workflows
In brief: OpenAI has rolled out "Computer History" for its ChatGPT desktop client on macOS, an event-based tracking feature that logs user activity to build an operational timeline. The tool enables ChatGPT and Codex to grasp workflow context, finish interrupted tasks, and recommend tailored automations, using an opt-in architecture that skips continuous screen captures in favor of lightweight event streams.
by Team Mocchi's
Human-AI interaction is rapidly moving past the constraints of isolated chat boxes. With the latest update to its macOS desktop application, OpenAI has rolled out a feature called Computer History, engineered to turn everyday computer actions into a continuous stream of operational context. The primary objective is to allow ChatGPT and the Codex code generation infrastructure to understand user workflows, propose proactive automations, and resume tasks left halfway through.
System events instead of continuous screenshots
The premise of Computer History inevitably draws comparisons to similar operating system-level memory features, most notably Microsoft's Windows Recall. However, OpenAI's architectural implementation diverges significantly in how user data is captured.
As reported by The Verge, the technology does not record video, audio, or periodic desktop screenshots. Instead, the application captures a structured stream of "events": keystroke sequences, mouse clicks, application focus changes, and file interactions. This data builds a searchable chronological timeline that the model can reference to understand the logical progression of work without having to parse heavy visual feeds continuously.
In a practical demonstration shared by OpenAI's developer experience team, the assistant identified the last document an engineer edited, verified whether it had been shared with colleagues on Slack, and generated a structured breakdown of the morning's activities.
Privacy controls and granular exclusion filters
Logging interactions at the desktop level naturally raises scrutiny regarding enterprise security and data privacy. OpenAI has implemented the capability on a strict opt-in basis, ensuring recording remains disabled until explicitly activated by the user.
From a control perspective, the feature includes exclusion filters for specific desktop applications and web domains, as well as the ability to purge individual entries from the recorded timeline. Ari Weinstein, Product and Engineering lead at OpenAI, noted that Computer History automatically ignores activity conducted within private or incognito browsing sessions.
The gathered event data serves a dual purpose: providing instant context when responding to ad-hoc user prompts, and identifying repetitive workflows so the model can suggest custom automated macros.
From passive chatbots to contextual operating agents
The arrival of this operational memory layer marks a strategic milestone for OpenAI's desktop ecosystem. Until now, delegating complex tasks to AI agents required users to manually articulate background parameters, upload relevant files, and explain previous steps.
By indexing system actions chronologically, the model gains cross-application awareness across development tools, internal messaging platforms, text editors, and browsers. This convergence positions ChatGPT not merely as an external advisor answering direct questions, but as an integrated background agent capable of bridging tasks across disparate software environments.
Mocchi's take
ChatGPT's shift toward event-based system tracking highlights where enterprise productivity is headed: from manual prompt engineering to contextual, background automation. For businesses and software engineering teams, relying on structured event streams rather than invasive screen captures represents a far more viable path for both computational efficiency and regulatory compliance, yet it demands strict governance over sensitive data and credential exposure. As software builders, we must design modern B2B architectures that natively expose structured context to system-level agents, preparing for workflows initiated directly by user activity rather than isolated chat prompts.