Chronicle

Definition: Chronicle is an AI-powered memory augmentation tool and screen-context engine developed by OpenAI for the Codex macOS application. Technically classified as a "Screen-Aware Memory Builder," it functions as an opt-in research preview that utilizes computer vision and Optical Character Recognition (OCR) to transform recent screen activity into structured Markdown memories for Large Language Models (LLMs).
Core Value Proposition: Chronicle exists to eliminate "context-switching fatigue" and the need for repetitive prompting. By bridging the gap between a user’s visual workspace and the Codex model’s context window, it allows the AI to maintain a continuous understanding of a user's workflow, tools, and project status without manual input. Key keywords include AI screen context, Codex memories, automated workflow summarization, and macOS developer productivity.

Screen-Based Contextual Awareness: Chronicle utilizes macOS Screen Recording and Accessibility permissions to capture visual data from the user’s desktop. Sandboxed background agents process these captures to identify active windows, code snippets, documentation, and communication threads. This allows Codex to "see" what the user is working on in real-time, providing a high-fidelity context that traditional text-only prompts cannot replicate.
Automated Memory Consolidation: The system periodically runs an ephemeral Codex session—often powered by models like GPT-5.4 or GPT-5.4-mini—to summarize recent activity. These summaries are converted into unencrypted Markdown files. This "memory building" process translates raw pixels and OCR text into semantic insights, which are then stored locally in the $CODEX_HOME/memories_extensions/chronicle/ directory.
Granular User Control and Privacy Management: Chronicle features a "Pause/Resume" toggle accessible via the macOS menu bar, allowing users to instantly stop screen capture during meetings or while handling sensitive data. Furthermore, the system is designed with an ephemeral storage architecture: raw screen captures are stored in a temporary directory ($TMPDIR) and are automatically purged after six hours, ensuring that only the synthesized memories persist.

Pain Point: Manual Context Restating: Developers and power users often waste significant time re-explaining project structures, tool configurations, or recent bug reports to an AI. Chronicle solves this by providing "zero-start" context, where Codex already knows the state of the user's IDE, browser tabs, and terminal.
Target Audience:

Software Engineers: Users navigating complex codebases who need the AI to understand cross-file dependencies and local environment settings.
Data Scientists: Professionals working between Jupyter Notebooks, Stack Overflow, and local documentation.
Power Users & ChatGPT Pro Subscribers: macOS users who rely on the OpenAI ecosystem for deep research, workflow automation, and technical assistance.

Cross-App Debugging: Using Codex to solve a bug by letting it see the error in a terminal, the relevant code in an IDE, and the API documentation in a browser simultaneously.
Automated Workflow Recall: Picking up a project after a break where Chronicle reminds Codex of the specific tools and libraries the user was previously utilizing.
Identifying Information Sources: Codex can use Chronicle memories to pinpoint exactly which Slack thread or Google Doc contains the information needed for a specific task.

Differentiation: Unlike traditional LLM "Memory" features that rely on past chat history, Chronicle is proactive and environment-aware. It captures context that was never explicitly typed into a chat box. While competitors often require manual file uploads or specific integrations, Chronicle observes the entire macOS environment to build a holistic memory of the user’s work.
Key Innovation: Local-First Semantic Storage: Chronicle’s approach to storing memories as local Markdown files is a significant innovation in AI privacy and transparency. Users can manually inspect, edit, or delete their memories at any time. This local-first architecture ensures that the "digital footprint" of the user’s workflow remains under their direct control while still being accessible to the AI for advanced reasoning and assistance.

How do I enable Chronicle on macOS? To enable Chronicle, you must be a ChatGPT Pro subscriber on macOS (outside the EU, UK, and Switzerland). Open the Codex app, navigate to Settings > Personalization, and ensure "Memories" is toggled on. Then, enable "Chronicle" and grant the required macOS Screen Recording and Accessibility permissions in System Settings.
Where are Chronicle memories and screenshots stored? Chronicle stores temporary screen captures in the $TMPDIR/chronicle/screen_recording/ folder, which are deleted after six hours. The consolidated memories (Markdown files) are stored locally in ~/.codex/memories_extensions/chronicle. These files are unencrypted, allowing users to review or modify the data the AI has "remembered."
Does Chronicle pose a security risk for prompt injection? Yes, using Chronicle increases the risk of "indirect prompt injection." If a user views a website or document containing malicious instructions, the background agents processing the screen might interpret those instructions as valid commands for Codex. Users should pause Chronicle when browsing untrusted content or viewing sensitive credentials.
Which models does Chronicle use to generate memories? By default, Chronicle uses the same model configured for your Codex instance. Users can specifically define a consolidation model, such as "gpt-5.4-mini," by updating the [memories] section of their Codex configuration file (consolidation_model = "gpt-5.4-mini").

Build Codex memories from recent screen context.