OpenAI Turns Mac History Into Agent Memory

OpenAI’s enterprise customer base reportedly grew 32% in July 2026. On August 14, CNBC attributed that figure to a source who said CFO Sarah Friar told investors that enterprise revenue now exceeded revenue from the ChatGPT-led consumer business. Separately, available product reporting described Computer History as an opt-in local timeline of selected Mac activity that ChatGPT and Codex could use, without establishing retention periods, the location of derived memories or the controls available to enterprise administrators.

Computer History moves memory beyond the chat

Computer History captures a local timeline across selected apps and sites. ChatGPT Work can recover context from that timeline, while ChatGPT and Codex can use the resulting memories. ChatGPT can therefore draw on activity outside a conversation even when the user never narrated it to the assistant.

OpenAI tested that mechanism before naming the product. On April 21, 2026, 9to5Mac reported that Chronicle, a macOS research preview for Pro subscribers, built Codex memories from recent screen captures. On June 4, OpenAI announced a ChatGPT memory summary page that lets users review and steer what ChatGPT remembers.

Chronicle gathered context; the summary page let users inspect memory. Computer History joins those steps. Its app-and-site selector governs what enters the timeline, but available reporting does not establish whether those selections also limit what agents retain as memory.

Agents need context when no one is prompting them

Technical work rarely lives in one application. A repository contains the current implementation, a terminal holds failed commands, a browser supplies documentation and a Slack message may contain the request that changed the requirement.

Cursor Automations makes the gap concrete. Cursor lets a codebase addition, a Slack message or a timer launch an agent. Because the user does not provide a fresh prompt at execution, Cursor must match the trigger to the correct repository, constraints and prior decisions.

An April 18 ZDNET report described OpenAI adding computer control, an in-app browser, automation memory and plugins to the Codex desktop app. Each tool gives Codex another place to act and another source of state to interpret. An agent that starts on a timer cannot rely on a person to supply the missing context.

Recall shows that deletion has two targets

Microsoft showed what happens when an operating-system vendor treats capture as retrieval before users accept its authority. The company delayed Recall after privacy and security criticism, then rebuilt it as an opt-in feature with an uninstall option. Microsoft also said Recall would encrypt its search-index database and require Windows Hello authentication and proof of presence before displaying the timeline.

Those controls addressed enablement, storage, authentication and removal. A 2025 assessment nevertheless found that Recall could be opened with a PIN after biometric setup and that sensitive-data filtering remained unreliable. Microsoft’s design showed that local storage and opt-in activation still depend on authentication and exclusion rules working as promised.

OpenAI builds Computer History on macOS rather than an operating system it controls. Available reporting describes the source timeline as local, but it does not locate the memories that ChatGPT or Codex derives from it. It also does not say whether clearing the timeline clears those memories, leaving an administrator with two possible deletion targets.

A managed fleet needs a policy owner

Consider a software company enabling Computer History for engineers on managed Macs. The company might permit capture in Xcode and internal documentation while excluding Slack, password managers and customer-support systems. The published description supports app and site selection; it does not say whether the employee or an administrator makes that selection.

If employees control the list, the company must decide whether individual choices satisfy its deployment policy. If administrators control it, buyers need to know how those settings are enforced across a fleet. Available reporting does not identify which model Computer History uses.

VentureBeat reported on April 24, 2024, that OpenAI had added administrative controls for API customers. Those controls do not establish who governs Computer History on managed Macs, and the available Computer History reporting does not connect them.

The same deployment needs an offboarding rule. A company must decide whether an administrator can export, preserve or delete a departing worker’s timeline and whether deleting that timeline also removes derived memories.

OpenAI’s enterprise customer base reportedly grew 32% in July, so more buyers now face that deployment decision. From the available reporting, they can assess the continuity benefit but cannot yet write the capture, retention and offboarding policy for a managed fleet.