ChatGPT now records how you work, then offers to automate it

The question in the hook is the right one to ask first: are you willing to let AI follow how you work all day in exchange for what it gives back?

Here is what is actually on the table, because the details changed this month and the change matters.

In April, OpenAI shipped a preview called Chronicle inside the Codex app. It took periodic screenshots of your screen, sent selected frames off to be processed, and built memories out of them. Screenshots sat on your machine for 6 hours before deletion. People compared it to Windows Recall, and the privacy criticism was loud.

This month OpenAI replaced Chronicle with Computer History, and the design is different in a way worth understanding. It does not capture your screen. It takes no screenshots, records no video, and never touches your microphone or system audio. Instead it reads interaction events through the macOS accessibility system: clicks, typing, keyboard shortcuts, app switches. Private browsing stays out entirely. Those events get turned into text summaries and saved on your own machine as Markdown files.

What you get back is a timeline in Settings, grouped by day and time, showing which apps fed each entry. You can open any memory file in Finder or delete it. ChatGPT and Codex can then answer questions about your own recent work: what you were doing before you stepped away, where that document you opened yesterday lives, a summary of yesterday for your standup.

The part with real consequence is what the timeline proposes. When it spots work you repeat, it offers to turn that workflow into a reusable Skill or a scheduled automation, and Codex builds it from what it recorded.

Now the costs, stated plainly. It is opt-in and macOS only. The memory files sit on your disk unencrypted for at least 48 hours, and OpenAI's own documentation says other programs running as your macOS user may be able to read them. It consumes rate limits quickly. And because recorded activity flows back into the model, it widens your exposure to prompt injection. One security write-up called it a keylogging and event capture system, and that description is fair.

One correction worth carrying into the video: Screen Context is a separate feature. It lives in ChatGPT Voice, you switch it on in Voice settings, and it grabs the frontmost window only when you ask it to take a look. That one is on demand. Computer History is the one that runs all day.

Follow for more:

  • https://www.instagram.com/ai.with.mo/

    • Course Registration: https://halaqa.app/enrollment?course=start-with-ai

    The trade, stated clearly

    Every productivity feature has a price, and this one names it clearly. To get a system that knows what you worked on, you hand it a record of what you did. The trade used to be worse. In April, Chronicle asked for your screen: periodic screenshots, selected frames processed on servers, memories built from images. The comparison to Windows Recall was immediate and the criticism stuck. Computer History, which replaced it this month, asks for less. It reads interaction events through macOS accessibility rather than looking at your display, which is a genuinely smaller ask, though it is still a record of your working day sitting on your disk. Whether that trade is worth it depends on what you actually do all day. Someone drafting client contracts and someone editing spreadsheets are exposing very different things.

    What it actually records

    The accuracy point most coverage gets wrong: Computer History does not watch your screen. OpenAI switched approaches. It captures interaction events exposed by macOS accessibility, meaning clicks, typing, keyboard shortcuts, and app switches. It takes no screenshots, records no video, and never touches your microphone or system audio, and anything you do in private browsing stays out. Those raw events are periodically compressed into text summaries and written to your machine as Markdown memory files. In Settings you get a timeline that groups those summaries by day and time and shows which apps contributed to each one, with an option to reveal the underlying file in Finder or delete it. From there ChatGPT and Codex can reference the memories in conversation. Ask what you were working on before lunch, where a document you opened yesterday lives, or for a summary of yesterday to bring to a standup, and the answer comes from your own recorded activity rather than from you retyping the context.

    It spots the repetition before you do

    Everything above is memory. The piece that changes how you work is what the timeline offers next. When the system finds work you repeat, the timeline entry proposes turning that workflow into a reusable Skill or a scheduled automation, and Codex builds it from the recorded steps. Read that sequence again, because it closes a loop OpenAI has been assembling for a while: saved memories, then chat history reference, then Chronicle, then Record and Replay, which turned a workflow you demonstrated into a reusable skill. With Computer History the system observes the work, remembers it, and proposes the automation without you identifying the opportunity. You are no longer the one who has to notice that you do the same eight steps every Monday morning. The practical consequence for anyone with a repetitive week is large, and it is also the reason the feature is worth being careful with. A system that can spot your patterns has, by definition, a detailed picture of your patterns.

    What to weigh before you switch it on

    Four things to weigh before you switch it on. It is opt-in and macOS only, so nobody enables it for you. The memory files are stored locally without encryption for at least 48 hours, and OpenAI's documentation states that other programs running as your macOS user may be able to access them, so the risk extends past OpenAI to anything else on your machine. It consumes rate limits quickly. And feeding recorded activity back into the model widens your exposure to prompt injection, where a malicious instruction sitting in something you viewed gets treated as an instruction. A security write-up described the whole thing as a keylogging and event capture system, which is uncharitable in tone and accurate in substance. If you decide to try it, the sensible approach is the same one you would use with any monitoring tool: turn it on for a defined stretch of ordinary work, look at what lands in the timeline, and check whether the automations it proposes are ones you actually want. Then decide. Also worth keeping straight for anyone confused by the naming: Screen Context is the separate Voice feature that captures your frontmost window only when you ask it to take a look, and organizations can disable it entirely.

    Prompt

    # COMPUTER HISTORY IN CHATGPT: WHAT IT IS AND HOW TO DECIDE
# Opt-in. macOS only. Replaced the Chronicle preview in August 2026.

# WHAT IT RECORDS
# Interaction events read through macOS accessibility:
#   clicks, typing, keyboard shortcuts, app switches
# Converted periodically into text summaries
# Saved locally as Markdown memory files

# WHAT IT DOES NOT RECORD
# Screenshots. Screen recordings. Microphone. System audio.
# Private browsing activity is excluded.

# HOW TO TURN IT ON
# 1. Open the ChatGPT or Codex app on macOS
# 2. Settings, then Personalization
# 3. Enable Memories, then enable Computer History
# 4. Grant Accessibility permission when macOS asks
# You can pause it at any time while working on something sensitive.

# WHAT YOU CAN ASK ONCE IT HAS DATA
"What was I working on before lunch?"
"Where does the document I opened yesterday live?"
"Summarize what I did yesterday so I can bring it to standup."
"Which apps did I spend the most time in this week?"

# THE PART THAT SAVES REAL TIME
# The timeline watches for work you repeat. When it finds a pattern,
# the entry offers to turn that workflow into a reusable Skill or a
# scheduled automation. Codex then builds it from the recorded steps.
# You do not have to notice the repetition yourself.

# HOW TO REVIEW AND DELETE
# Settings shows a timeline grouped by day and time, with the apps
# that contributed to each entry. Any entry can be opened in Finder
# or deleted. The files are plain Markdown, so you can read them.

# THE RISKS, PLAINLY
# 1. Memory files are stored unencrypted for at least 48 hours.
#    OpenAI documents that other programs running as your macOS user
#    may be able to access them.
# 2. It consumes rate limits quickly.
# 3. Recorded activity flowing back into the model raises your exposure
#    to prompt injection from content you viewed.
# 4. It is a record of your working day. Treat it accordingly.

# A SENSIBLE WAY TO TRY IT
# Turn it on for one ordinary work week. Read the timeline at the end.
# Check whether the automations it proposes are ones you actually want.
# Pause it whenever you are working on anything confidential.
# Then decide.

# DO NOT CONFUSE IT WITH SCREEN CONTEXT
# Screen Context is a different feature, and it lives in ChatGPT Voice.
# You enable it in Voice settings and it captures the frontmost window
# only when you ask ChatGPT to take a look. It runs on demand.
# Organizations can disable screen context entirely.