OpenAI’s Computer History Turns Mac Activity Into ChatGPT Memory—Without Screenshots

OpenAI’s Computer History gives the ChatGPT macOS app a memory of activity outside the chat window. Announced on August 13, 2026, the opt-in feature uses interaction events from selected apps and websites to build a searchable timeline and local memories that ChatGPT and Codex can reference later.

The design is less visually invasive than screenshot-based activity tracking, but it is not risk-free: accessibility events can include typed text, and the resulting memory files are stored locally as unencrypted Markdown.

How Computer History works

After a user grants macOS Accessibility permission and selects the apps or websites to include, Computer History records interaction events exposed through macOS. These can include clicks, typing or keystrokes, keyboard shortcuts, app switching and related application or website context.

The event stream is then used to create a timeline and text memories. In practice, ChatGPT or Codex can use that context to help answer questions such as what a user was working on, where a task stopped or which steps were used in an earlier workflow.

The feature is off by default and remains limited to the apps and sites a user allows. Private or incognito browser activity is excluded.

What it does not capture

OpenAI’s description says Computer History does not record:

• Screenshots
• Screen recordings
• Microphone input
• System audio
• Private-browser activity

That makes it different from a system that continuously saves images of the screen. However, “no screenshots” does not mean “no sensitive information.” If a user types confidential material into an included app or website, the interaction record may still contain meaningful text or context.

Computer History vs. Chronicle and Microsoft Recall

Chronicle used screenshots

Computer History replaces the earlier Chronicle research preview, which built context from periodic screen captures. The newer feature uses an accessibility-event stream instead, so it does not need to capture screen pixels or audio to create a history.

That is a narrower form of capture, but not necessarily a negligible one. A sequence of keystrokes, clicks and app switches can still reveal what someone was doing, especially when it is summarized into a durable memory.

Microsoft Recall is snapshot-based

Microsoft Recall is primarily a screenshot or snapshot-based activity-history system for Copilot+ PCs. Computer History pursues a similar goal—giving an assistant searchable context about past computer activity—but relies on interaction events and text memories rather than a visual snapshot archive. Reporting also describes Computer History as sending event data to OpenAI for processing, rather than keeping the entire workflow local.

The practical distinction is important: Computer History avoids storing visible passwords, documents, conversations or images as screen pixels, but its event stream can still expose typed material and workflow context.

What is stored for 48 hours—and what remains afterward?

The data lifecycle has two different parts:

1. Temporary interaction data: Raw event files are stored on the Mac for up to 48 hours and then removed as part of the processing lifecycle, according to the supplied reporting.
2. Generated memories: OpenAI processes the events server-side to turn them into usable memories. The resulting memories are stored locally as plain-text Markdown files.

The persistent files are readable and editable as text rather than being locked inside an encrypted database. One report identifies a Codex memories directory under $CODEX_HOME/memories/extensions/skysight/, but the supplied evidence does not establish that path as an official, universal location for every installation.

Deleting Computer History or its history should remove associated local memories and stop future collection. The available feature-specific material does not provide a precise server-side deletion timetable, so deleting the local history should not be treated as proof that every copy involved in processing disappears instantly. Ordinary legal, security or fraud-prevention obligations may also affect deletion.

The main privacy and security risk: unencrypted local files

OpenAI’s documentation warns that Computer History files may contain sensitive information and are not encrypted by Computer History. Other programs running as the same macOS user may be able to access them.

That creates a different risk profile from ordinary browser history. A memory file could consolidate project names, workflow steps, app usage, typed material and account-related context into a readable text corpus. Malware, an infostealer, a remote-access tool or another process running under the logged-in account could potentially read and exfiltrate that information.

Full-disk encryption can help protect a powered-off Mac, but it does not normally stop software running inside an unlocked user account from accessing files that account can read. The feature’s local storage therefore matters even though the raw event files are temporary.

Prompt injection is a possible downstream risk

Computer History can also create a prompt-injection pathway. Untrusted instructions displayed in a website or entered into an included app could be summarized into a memory. If ChatGPT or Codex later treats that memory as trusted context, the text could attempt to influence a future response or action. This is a security inference, not evidence that every such attack will succeed, but it is a reason to treat recorded external content as untrusted.

Who can use it?

Computer History is rolling out on macOS for ChatGPT Pro, Business and Enterprise users. At launch, it was unavailable in the European Economic Area, Switzerland and the United Kingdom, with access for those regions indicated for later.

Access works differently by plan:

• Pro: Users can enable the feature themselves.
• Business: A workspace administrator must first allow access; each member must still opt in individually.
• Enterprise: Administrators can grant access by role, but users still choose whether to enable it. Enterprise users may also need a sufficiently recent desktop-app version for the controls to take effect.

The supplied launch material does not establish a separate Computer History-specific model-training rule for every plan. Business and Enterprise users should follow their workspace’s existing data controls, while Pro users should check their account’s model-improvement settings rather than assume that Computer History is categorically excluded.

Controls users and administrators get

Users can:

• Opt in or leave the feature disabled
• Pause or turn off collection
• Choose which apps and websites contribute
• Review timeline items
• Delete recorded history and memories

Business and Enterprise administrators can control whether members are allowed to access the feature in the first place. That two-step model—administrator permission followed by individual consent—prevents an organization from silently activating Computer History for everyone.

For a workplace deployment, it would be prudent to exclude password managers, financial, health and legal systems, privileged administration consoles, confidential source-control or incident-response tools, and communication apps containing other people’s private information. Those exclusions are operational guidance, not a guarantee that every sensitive value will be filtered automatically.

Could the history create skills or automations?

The feature’s most useful long-term possibility is procedural memory. A timeline showing a successful sequence of actions could help Codex reconstruct a repeatable workflow as a checklist, script or reusable skill.

It could also provide context for a future request such as reproducing a report-preparation process used earlier in the week. But that is a potential workflow built on recorded context—not evidence that Computer History itself automatically creates or runs scheduled automations today.

Any automation built from this kind of history should require explicit approval, constrained permissions, reviewable steps and safeguards against instructions injected through websites, documents or messages.

The practical verdict

Computer History is best understood as event-based AI memory, not screenshot surveillance. It reduces the amount of visual and audio data captured compared with Chronicle or snapshot-based systems such as Recall, while still giving ChatGPT and Codex a detailed record of selected computer activity.

That trade-off may be useful for users who frequently resume complex workflows, but the unencrypted local Markdown memories are a serious consideration. Anyone enabling the feature should use narrow app and website allowlists, exclude high-risk systems, pause collection when handling sensitive material and regularly inspect or delete the resulting memories.