PRODUCT
In the early hours, OpenAI launched a new feature for Codex called "Chronicle," enabling the AI to work based on screen content.
Codex Gains a New Memory Source: Your Screen
From now on, when you speak to Codex, you no longer need to repeatedly explain what "this" or "that" refers to. It knows what you are currently looking at, what error just occurred, and what project you were working on two weeks ago.
Chronicle is an extension of last week's "memories" preview. While the source of "memories" was conversation history, Chronicle adds a layer of screen context.
It is currently available only to ChatGPT Pro users on macOS. It is not yet supported in the EU, UK, or Switzerland, and its status is an opt-in research preview.
Scenario 1: Direct Screen Debugging
The first of three demonstrations in the official video is called "Use what's on screen."
On the left side of the screen is a CI failure log. The user opens Codex and asks just one question: "why is this failing?"
Codex did not ask the user to paste the error message, nor did it ask which repository it was. Through Chronicle, it captured the context on the screen and spent 1 minute and 43 seconds investigating on its own.
Codex looked at the CI log on the screen and located the specific job, file, and line number.
It ultimately pinpointed the `build-preview` job in GitHub Actions, specifically at line 2 of `src/pages/preview-build-fixture.astro`:
astro
const articleCount: string = 404;
The conclusion was a TS2322 type error: 404 is a number but was declared as a string. Codex also noticed that the PR title read "exercise CI failure handling," indicating this was an intentionally inserted type error fixture.
The bottom of the response cited 1 memory reference, using the repo-specific validation loop context from line 2122 of MEMORY.md as a reference prior to local reproduction.
Official materials also included a comparison screenshot with Chronicle turned off. For the same question "why is this failing," Codex simply replied, "I do not know what this refers to yet":
When Chronicle is off, Codex asks for the specific error text or link.
Scenario 2: Screen Context Completes "This" and "That"
The second scenario is called "Fill in missing context."
The user inputs: "Sync with the latest docs draft changes and message Romain."
The user did not specify which draft is the "latest docs draft" or who "Romain" is.
Codex parsed which draft and which Romain via Chronicle memory, then synchronized the file and sent a Slack message.
Through Chronicle, Codex found the document draft the user was editing, synchronized `chronicle.mdx`, and then called the Slack skill to send a direct message to Romain, informing him of the sync status and build results.
If you look closely at the process, you will see GPT thinking about who Roman is based on Chronicle, and the internal reasoning of memory parsing is written more explicitly: Memory points to the Google Doc Chronicle docs draft and Romain Huet on Slack.
Chronicle's memory recorded the most recent draft and Romain Huet's Slack ID.
Throughout the entire process, the user never clarified the words "this," "that," or "latest"; Codex filled them in itself.
Scenario 3: Remembering Your Tools and Workflows
The third scenario is called "Remember tools and workflows."
The user says: "Create an empty draft doc for the Chronicle launch copy to share with the team."
Codex checked memory, saw that the user habitually uses Google Drive for drafts, and directly created a Google Doc.
Codex first checked memory, confirmed the user's habit of placing drafts in Google Drive, and then directly called the Google Drive tool to create the file Chronicle Launch Copy [DRAFT]. A permission confirmation dialog appeared during the process.
In the comparison version with Chronicle turned off, Codex would ask what format you want: "What format should this be? I can draft it here, create a Google Doc, or prepare a Slack post."
Without Chronicle's memory, Codex does not know where to place launch communications.
OpenAI's own explanation is: Codex first uses Chronicle to locate the source; when it's time to actually work, it then reads the specific files, Slack threads, Google Docs, dashboards, or PRs.
How to Enable
Besides the aforementioned Pro and macOS restrictions, you must grant Screen Recording and Accessibility permissions.
Enablement path:
01 Open Codex Settings.
02 Go to Personalization and confirm Memories is turned on.
03 Open Chronicle located under Memories.
04 Click Continue on the confirmation dialog.
05 When macOS pops up the permission request, grant Screen Recording and Accessibility.
You can Pause or Resume at any time via the menu bar icon. Pause before meetings or when viewing sensitive content.
Technical Details
Screen recordings are stored in CODEX_HOME/memories_extensions/chronicle/, which defaults to ~/.codex/memories_extensions/chronicle/.
The memory itself is an unencrypted markdown file; users can read, modify, or delete it. OpenAI advises against manually adding new entries, but local modifications and deletions are supported.
The model used to generate memory defaults to the same model used by Codex. To change it, you can set it in config.toml:
toml
[memories]
consolidation_model = "gpt-5.4-mini"
Codex does not generate memory immediately; it skips active sessions, filters out sensitive information like keys, and writes in the background after the thread has been idle for a period.
Risks and Costs
The official documentation lists three risks:
First, rate limits. The sandboxed agent running in the background for Chronicle continuously consumes quotas; OpenAI's exact words are "uses rate limits quickly."
Second, prompt injection. If a webpage with malicious agent instructions appears on the screen, Codex might execute the instructions on the screen.
Third, data visibility. Memory files are unencrypted, meaning other apps on the same computer can also read these files.
Chronicle does not access the microphone or system audio; it only takes screen recordings. OpenAI specifically reminds users not to use Chronicle to record meetings without obtaining others' consent.
When Chronicle generates memory, the selected screen recording, OCR-extracted text, timestamps, and local file paths are sent to OpenAI servers for processing. After processing, the screenshots are not retained and are not used for training.
Why Is This Just a Research Preview Today?
The official blog states: "going forward, OpenAI is developing Codex into a more capable tool for builders beyond software engineers, and Chronicle is one step toward that goal."
The official statement on X was: "while we learn where it helps most and improve the experience."
Therefore, today's version is opt-in; Pro users who want to try it must enable it themselves in Settings.
Reference Materials
→ Chronicle Official Documentation: developers.openai.com/codex/memories/chronicle
→ Memories Official Documentation: developers.openai.com/codex/memories
→ Codex Official Blog: openai.com/index/codex-for-almost-everything
→ OpenAI Developers X: x.com/OpenAIDevs