ChatGPT's memory just moved from the chat window to the endpoint

OpenAI launched Computer History on Thursday, an opt-in ChatGPT feature that remembers a user's activity across apps and websites on a Mac and draws on that record in future conversations ^[1]. It is rolling out globally through the ChatGPT desktop app to Pro, Business, and Enterprise subscribers ^[2], which means the approval question for IT owners is no longer about a chat product but about a process that observes the desktop.

The boundary is the story. ChatGPT already retained preferences and personal details through saved memories and prior conversations when those settings were enabled ^[3]. Computer History widens the input to work performed across desktop software and websites, rather than waiting for a user to type or upload it ^[4]. OpenAI describes it as an extension of Chronicle, an earlier research preview that used recent screen context to create memories, and says the new implementation reduces token usage and adds privacy controls ^[5]. The announcement did not quantify the token reduction or describe the storage architecture behind the activity record ^[6].

That last omission is the one that matters for a security review. Screen activity can contain credentials, private messages, customer information, health data, financial records, and confidential company material, as runtimewire.com notes in its account of the launch ^[7]. The controls OpenAI has described are user-side: the feature is opt-in rather than on by default, and users can review activity through a timeline and manage the feature from the timeline or the menu bar ^[8]. Nothing in the announcement described administrator-level workspace enforcement, and OpenAI has not said which applications receive deeper integrations, how far back the timeline reaches, or whether retention limits differ by subscription or workspace ^[9]. So the enrollment decision currently sits with the employee, while the retention consequence sits with the company ^[10].

It also does not sit alone on the endpoint. In July, OpenAI combined Chat, Work, and Codex into a single macOS and Windows client, where Work can use local files and desktop apps with permission and Codex handles software-development tasks ^[11]. Computer History supplies persistent context to that same client, which reduces how often a task needs a fresh briefing ^[12]. Read as a system, a permissioned local-file agent and a continuous activity record are two halves of the same data path.

Rollout mechanics give governance teams a short window. Availability in the European Economic Area, the United Kingdom, and Switzerland will follow in the coming weeks ^[13], so a workspace with staff on both sides of that line will behave inconsistently for a period ^[14]. The launch applies to Mac even though the desktop app supports both macOS and Windows, and OpenAI gave no Windows date ^[15], which means the population in scope today is whatever fraction of your fleet runs macOS ^[16]. Plus, Free, Go, and Education users are excluded at launch ^[17] - the included tiers are the ones where people use ChatGPT all day ^[18].

Practical framing: this is a new endpoint data flow with a per-user switch, enabled from Settings > Integrations ^[19], not a feature to tick off in a vendor questionnaire.