ChatGPT's Computer History: AI That Watches Your Workflow and Writes Automations

OpenAI has rolled out Computer History for ChatGPT Pro, Business, and Enterprise subscribers on the macOS desktop app. Unlike the earlier Chronicle research preview that relied on screenshots, Computer History collects structured interaction events—clicks, keystrokes, shortcuts, and app switches—through macOS Accessibility APIs, plus the text and context those APIs expose. It explicitly does not capture screenshots, microphone audio, system audio, or private-browsing activity.

How the pipeline works

Raw interaction events are written to temporary files inside the ChatGPT App Group container on the local Mac. Other apps cannot read them without explicit authorization.

These temporary files are retained for a maximum of 48 hours before ChatGPT/Codex deletes them.

Periodically, a disposable Codex session is spawned; the event stream is fed to it, and Codex writes a plain-text Markdown summary. OpenAI states the server-side processing does not retain the events or use them for training unless legally compelled.

The resulting memory files are stored locally at $CODEX_HOME/memories/extensions/skysight/ (typically ~/.codex/memories/extensions/skysight/). They remain on disk until the user deletes them.

You can inspect the memories with a simple command: ls -la ~/.codex/memories/extensions/skysight/ Each file is readable with cat; unwanted entries can be removed manually.

Comparison with Chronicle and Windows Recall

The author includes a diagram contrasting three approaches:

Chronicle – screenshot-based capture.

Windows Recall – also screenshot-based.

Computer History – event-stream + Accessibility text, no pixels.

The author argues that avoiding screenshots dramatically reduces the security-audit surface because event streams are readable, trimmable, and explainable. However, they caution that Accessibility APIs can see every character typed into input fields, so the coverage is broader than many users expect.

User controls

Menu-bar icon: shows what is being captured, offers Pause / Resume , and can delete the most recent session for a specific app.

Settings → History page: reveals the local memory file in Finder; supports deletion of single entries or bulk cleanup for the last 10 minutes, 1 hour, 1 day, or all history. Deletion also removes the underlying interaction events and is irreversible.

Sensitive apps (e.g., messaging, banking) can be added to an exclusion list so their activity never enters the stream.

Core use cases

Resume context – ask “What was I doing before my break?” and get an instant summary instead of reconstructing window state manually.

Retrieve artifacts – describe a past artifact in your own words (“the proposal I looked at this morning”) and the system follows the timeline to the source document or conversation.

Automate repetition – when the timeline detects a repeated workflow, it suggests creating a skill or automation ; one click generates the code from the recorded steps.

The author emphasizes that Computer History hands over leads (timeline entries) rather than full copies; ChatGPT/Codex then fetches the original content when needed.

Risk and cost checklist

Prompt-injection surface expands – malicious web pages or app content could inject instructions that the model later executes (explicitly noted in OpenAI’s documentation).

Others’ privacy – the feature should be paused during conversations with other people unless they have given explicit consent; health, finance, and other sensitive apps should be excluded.

Chat context may still reach the model – while event files are not used for training, memories brought into a conversation become part of that conversation’s context, and whether that conversation improves the model depends on the user’s separate “Model Improvement” setting. The author recommends turning Model Improvement off.

Token consumption – summarization and memory generation consume tokens; OpenAI documents this under a dedicated “Token usage” section.

Outlook

The author views Computer History as a pivotal shift: the agent’s context source moves from “what you paste me” to “everything happening on your machine.” They expect every major AI harness to adopt a similar architecture, but remain uneasy about privacy given industry attitudes toward data-for-convenience trade-offs.