OpenAI has rolled out Computer History, an activity tracking system that compiles a timeline of operating system events to supply context directly to ChatGPT and Codex. The utility replaces an earlier internal project, Chronicle, marking a clean pivot away from multimodal screen capture. Instead of running heavy visual parsing pipelines over desktop frames, Computer History taps directly into macOS accessibility APIs to record clicks, keystrokes, shortcuts, and application switches at the OS level.
This event-stream architecture cuts the processing and bandwidth overhead of constant optical character recognition and video ingestion. The pipeline ignores raw screen recordings, microphone input, and system audio entirely, while blacklisting private browsing sessions by design. OpenAI transfers these raw system events to its servers for processing, translates them into structured text summaries, and writes them back to the user's local disk as plaintext Markdown files that map application usage and task timelines.
Transforming Routine Activity into Codex Skills
Computer History serves as an ingestion engine for procedural automation rather than a passive desktop diary. By monitoring recurring operational patterns across applications, the system identifies repetitive workflows and translates them into executable automations termed Skills.
When ChatGPT detects repeatable patterns in a user's activity, the timeline suggests a "Skill" or automation.
Instead of forcing engineers to manually draft specifications, users can prompt Codex to compile these captured interaction chains directly into reusable automation templates. The approach turns ambient daily habits into executable code, giving OpenAI a direct route to automating routine white-collar tasks from the ground up.
Security Architecture and Corporate Risk
Enterprise deployment introduces friction across corporate compliance, data sovereignty, and security boundaries. For Business and Enterprise accounts, the rollout mandates explicit administrator enablement, individual employee opt-in, and active Memories configuration. Users get menu-bar toggles to pause logging or configure blocklists for internal tools and domains. Regulatory constraints also leave the feature entirely unavailable across the European Economic Area, Switzerland, and the United Kingdom.
Local storage mechanics present immediate attack surfaces. OpenAI holds raw event logs in its macOS application container for up to 48 hours, but generated Markdown memory summaries persist on the local drive in unencrypted plaintext until manual deletion. Any background process running with the same local user permissions can read these records. While OpenAI states that processed context is purged from remote servers and withheld from base model training, active chat sessions that pull these summaries remain subject to tenant data policies. Moreover, prompt injection via maliciously crafted web pages poses a direct threat to the parsing engine, prompting OpenAI to advise enterprise teams to explicitly exclude sensitive communication, health, and financial utilities.
Security teams face a sharp tradeoff: granting system-level telemetry and tolerating unencrypted local audit trails, or opting out and foregoing autonomous workflow synthesis entirely.