OpenAI’s Computer History turns ChatGPT into a form of working memory for activity on your Mac. The feature matters to anyone who regularly uses Artificial Intelligence (AI) for documents, research, or recurring office tasks. It promises less searching and repetition, but records a much broader part of the workday than an individual chat does.
More context for ongoing work
According to OpenAI, Computer History captures interaction events across apps and websites. These include clicks, keystrokes, keyboard shortcuts, and switches between apps, read through the macOS accessibility system. That system gives apps selected ways to observe or control parts of the user interface.
The feature replaces an earlier research preview called Chronicle, which relied on screenshots. OpenAI says Computer History does not capture screenshots, screen recordings, microphone input, or system audio. Activity in private browsing mode is also excluded. Those boundaries reduce data collection, although a record of typing, clicks, and app changes can still reveal a great deal about someone’s work.
The events are periodically converted into text summaries and stored locally as Memory files in Markdown, a plain-text format that uses simple markers for headings and lists. A timeline arranges the summaries by date and time and identifies the apps involved. As the description of Computer History explains, ChatGPT and Codex can use these records as context.
One practical example is finding a document you edited recently. Instead of remembering its filename and folder, you could ask ChatGPT to identify the relevant source from your recorded activity. Another example is producing stand-up summaries, meaning short reports on what a team member has recently worked on. Computer History is meant to assemble the relevant actions from the timeline.
The system can also identify recurring workflows and suggest an automation, according to OpenAI. It presents this as a Skill, meaning a reusable template based on an observed process. You can then ask Codex to turn the workflow into an automation. This is most useful when a task follows a stable pattern; the benefit is smaller for rare or constantly changing work.
Control remains local and layered
Computer History is an opt-in feature for the Mac. In Business and Enterprise workspaces, an administrator must first make it available. Individual users must still consent, and ChatGPT’s Memories feature must also be enabled.
Include and exclude lists determine which apps and websites may contribute data. Recording can also be paused and resumed from the macOS menu bar. That provides more control than blanket monitoring, but it requires careful setup: if you fail to exclude sensitive human resources, customer, or financial apps, their use may become part of the local work history.
Local storage is an advantage, but it is not a complete privacy guarantee. The files contain condensed information about work patterns and can be used as context by ChatGPT or Codex. The supplied reports do not explain in detail which parts leave the computer in every processing situation or how long individual contents are retained. Those gaps matter particularly to organizations with internal rules for confidential information.
For Switzerland, the current position is clear: OpenAI says Computer History is not available in Switzerland, the European Economic Area, or the United Kingdom. Swiss users therefore cannot test it in normal use at present. The sources do not provide a date for wider availability.
Pros and Cons of Computer History
Pros:
- Less repetition – ChatGPT can use previous work as context instead of starting from scratch in every session.
- Faster retrieval – Recorded activity may identify recently used documents even when you have forgotten their filenames.
- Automatable routines – Recurring workflows can be recognized as Skills and converted into reusable templates.
- Layered consent – Administrative approval, individual consent, Memories, and exclusion lists provide several levels of control.
Cons:
- Detailed work profile – Clicks, typing, app changes, and timestamps collectively reveal a substantial picture of your activity.
- Sensitive company data – Poorly configured include lists may pull confidential work contexts into the summaries.
- Manipulable context – Documents and websites can contain hidden instructions intended to mislead an AI system.
- Limited availability – The Mac feature is currently unavailable in Switzerland and several other regions.
A practical way to begin
If you are a beginner, starting with your entire digital workday would be unwise. A more manageable first use would be one low-sensitivity workflow, such as working with public research material or finding general project documents. This lets you test whether the summaries genuinely save time without immediately including personnel records, confidential messages, or login credentials.
Step 1: Limit collection
- Select only the apps and websites required for the workflow you want to test.
- Exclude tools containing confidential company, personnel, or financial information.
- Pause recording whenever you move temporarily beyond the defined workflow.
Step 2: Review the results
- Regularly inspect the generated summaries for unexpected or sensitive content.
- Compare the time saved on retrieval and summaries with the extra review work.
- Expand collection only after the value of the limited test is clear.
More advanced users can focus on highly repeatable processes and turn suitable patterns into Skills. Each suggested workflow should still be reviewed before reuse, especially when external documents or websites are involved. Automation multiplies useful steps, but it can also multiply poorly selected context.
A case in Connecticut demonstrates the problem in unusually direct form. A plaintiff placed tiny white text on a white background in court filings to influence a possible AI review in his favor. This technique is called prompt injection, meaning a manipulative instruction embedded in content processed by an AI system. Unusual blank space exposed the attempt, and the judge objected to the hidden commands. The case did not involve Computer History, but it shows why automatically read material should not be treated as neutral data.
Costs, speed, and unresolved security issues
The sources do not identify a separate price for Computer History. At the same time, AI costs are increasingly divided between cheaper models and more expensive speed tiers. According to a report on competition among OpenAI, Anthropic, and Chinese providers, OpenAI cut the price of GPT-5.6 Luna by 80 percent, while Anthropic offered Claude Opus 5 at half the price of Fable 5.
The token price index cited in that report, which tracks the cost of units processed by language models, fell by almost a quarter for leading US providers from mid-July. Some enterprise customers are also moving from flat subscriptions to usage-based billing. More automatically supplied context may therefore affect the volume of processed data as well as convenience, depending on the plan, although the sources give no specific cost calculation for Computer History.
OpenAI is also previewing an Ultrafast mode for GPT-5.6 Sol. According to the provider, it can produce up to 750 output tokens per second and run up to 14 times faster than standard processing. It initially operates through the application programming interface for selected customers and is expected by the report to cost more; the existing Fast Mode already costs about twice as much as standard processing. The listed uses for the faster mode include customer support, e-commerce, and analysis during an active system outage. Speed, however, does not correct unsuitable context or solve privacy concerns.
A separate report illustrates why stored AI context can become a security target. Researchers reportedly decrypted internal reasoning data, meaning information generated while a model processes a request. Their review of more than 300,000 records from public source-code repositories and log files exposed personal identifiers, credentials, and plain-text passwords. The researchers also warned that specially prepared reasoning data could carry hidden prompt injections.
This weakness involving internal reasoning data is not evidence of a vulnerability in Computer History. It does show that context files, logs, and cached model information may contain sensitive material when handled carelessly or protected inadequately. Local storage is therefore not automatically risk-free, just as faster processing is not automatically better.
Computer History could increase ChatGPT’s practical value by reducing repeated explanations and recognizing recurring work. Against that benefit stand a detailed activity record, manipulable input, and unanswered questions about processing and cost. For Swiss users, the trade-off remains theoretical for now; the unresolved risk is that convenience may expand faster than control over the recorded context.
Sources
- ChatGPT’s New Computer History Turns the Workday Into AI Context – Unknown, 2026-08-14
- ChatGPT Gets a Memory: What Is Behind Computer History – Unknown, 2026-08-14
- 14 Times Faster: OpenAI Launches Ultrafast Mode for GPT-5.6 Sol – Unknown, 2026-08-14
- Researchers Decrypt Internal Reasoning Data From AI Models – Unknown, 2026-08-14
- US Plaintiff Hides Manipulative AI Commands in Court Documents – Unknown, 2026-08-15
- OpenAI and Anthropic in Price War as Chinese AI Rivals Gain Ground – Unknown, 2026-08-14
