OpenAI ChatGPT Computer History feature launch

Computer History is not a screenshot tool. It captures interaction events - clicks, typing, keyboard shortcuts, and app switches - exposed through macOS's accessibility APIs, then converts them into text summaries rather than storing images, video, or audio ^[1]. That is a substantive rebuild, not a rebrand, of the Chronicle research preview that used screenshots to build ChatGPT's memory of a user's activity ^[2]. The feature depends on ChatGPT's Memories system and only runs inside the macOS desktop app - it is unavailable through the API or Amazon Bedrock, which keeps it tied to interactive, individual use rather than programmatic access ^[1]. Raw event files live on the Mac for a maximum of 48 hours before OpenAI processes them server-side into summaries and discards them, while the resulting memory files are stored locally as plaintext Markdown until a user deletes them; OpenAI says none of it feeds model training ^[3]. When the system notices a workflow repeating, ChatGPT or Codex can offer to turn it into a reusable skill or scheduled automation ^[4].

OpenAI built Computer History in the shadow of Microsoft's Recall, the Windows 11 feature that screenshots a user's screen roughly every five seconds and generated substantial privacy backlash ^[6]. Press coverage reads Computer History as a direct answer to that backlash: no screenshots, off by default, and event logs a user can pause or purge in bulk ^[5]. The Register goes further, casting the whole design as OpenAI's attempt to avoid Recall's mistakes while building functionally the same capability - a persistent record of what a user did on their machine - through keystroke and click logging instead of images ^[3].

OpenAI's own documentation concedes two risks it built into the feature rather than one it merely inherited. First, Computer History raises the odds of prompt injection: if a user visits a site or app carrying hidden malicious instructions, that content can end up in the activity timeline and later be followed by ChatGPT or Codex as if it were the user's own intent ^[3]. Second, neither the raw event files nor the generated memory files are encrypted at rest, meaning any other program running under the same macOS user account could potentially read them ^[3]. On top of the technical exposure, OpenAI is withholding the feature from the EEA, UK, and Switzerland at launch, with access promised only for 'coming weeks' - a delay that tracks those regions' stricter data-protection scrutiny of persistent activity-tracking tools ^[7].

Reaction has split cleanly along how people use ChatGPT, and along who is doing the talking. The launch itself was entirely OpenAI-driven - the announcement and its main explainer video both came from OpenAI's own X and YouTube channels, with no independent technical review yet published elsewhere. On general-audience threads, the dominant read pushes back on that framing: commenters suspect the feature is a way for OpenAI to harvest fresh training data now that public web text is running out, and one widely-shared take reframes the real concern as an operating-system problem - unrestricted accessibility-API access is a broader security overreach, not just a privacy tradeoff. A more optimistic minority views it as an early, deliberately incremental step toward a genuine AI operating system. Developer-focused communities argue over something more concrete: engineers who rely on Codex note that Computer History undercuts the habit of starting a fresh, low-context chat, since the timeline keeps resurfacing unrelated prior projects. Practical mitigation threads have emerged in response, recommending users scope the feature to specific tools - IDEs, terminals, code hosts, documentation - while explicitly excluding password managers, banking apps, health portals, and private messaging.