Chronicle, an opt-in feature for ChatGPT Pro subscribers on macOS, augments Codex memories with real-time screen context to reduce repetitive prompting—but requires careful consideration of privacy trade-offs and geographic restrictions.
OpenAI has introduced Chronicle, a new research preview feature for its Codex coding assistant that aims to minimize the need for users to manually restate context during interactions. Available exclusively to ChatGPT Pro subscribers running the Codex app on macOS (and currently blocked in the EU, UK, and Switzerland), Chronicle captures periodic screen snapshots to enrich Codex’s memory system with visual and workflow context from the user’s active session.
The core premise addresses a common friction point in AI-assisted development: repeatedly explaining your current file, open tabs, or recent actions. By analyzing screen content—such as the code editor, documentation, or terminal output—Chronicle helps Codex infer relevant context automatically. For instance, if you’re debugging a specific function in VS Code, Chronicle might surface that file’s recent changes or related terminal commands without explicit prompting. Over time, it also learns your preferred tools and workflows (e.g., which linter you use for Python projects), reducing the need to specify "use pytest" or "run the dev server" in every thread.
Technically, Chronicle operates through sandboxed background agents that process screen captures. These agents extract OCR text, timing data, and file paths from temporary screenshots (stored briefly in $TMPDIR/chronicle/screen_recording/), then use a Codex session—typically powered by your default model or a specified consolidation_model like gpt-5.4-mini—to summarize activity into markdown memories. These memories reside locally under $CODEX_HOME/memories_extensions/chronicle/ (usually ~/.codex/memories_extensions/chronicle/), alongside the ephemeral screen captures. Notably, the memories are stored unencrypted, mirroring the handling of standard Codex memories, which means any sensitive information visible on your screen during capture could persist in readable form.
This local storage approach carries significant implications. OpenAI emphasizes that screen captures are deleted after six hours and never retained on its servers post-processing (unless legally required), but the unencrypted local files introduce risks. If malicious content appears on your screen—such as a webpage with hidden prompt injection instructions—Codex could potentially act on it when synthesizing memories. Users are advised to pause Chronicle during meetings or when viewing sensitive data, accessible via the Codex menu bar icon. Setup requires granting macOS Screen Recording and Accessibility permissions, with troubleshooting steps available if initialization fails due to restricted access.
The feature’s availability highlights OpenAI’s staged rollout strategy: limiting access to Pro subscribers on a single platform while gathering real-world usage data. For developers weighing the convenience against the privacy considerations, Chronicle represents an explicit trade-off—trading reduced cognitive load for heightened vigilance over screen content. As a research preview, it invites user feedback to refine both the utility and safeguards of context-aware AI assistance, particularly as similar techniques may evolve across other tools in the ecosystem.
For full details on enabling Chronicle, managing memories, and understanding the data flow, refer to the OpenAI Developers documentation.
Comments
Please log in or register to join the discussion