OpenAI has recently introduced a major feature called "Chronicle" for its Codex application. This feature records users' screen activities in the background, building a comprehensive "memory library" for the AI. This means that Codex can have a perspective similar to that of a human colleague, real-time understanding of the specific projects users are working on, the tools they are using, and their current progress. In subsequent interactions, users no longer need to repeatedly provide tedious preface explanations, as the AI can accurately understand their intent based on this "digital memory."

In terms of technical implementation, Chronicle demonstrates a high level of automation. Once activated, this feature runs silently in the background, with an internal agent performing deep analysis and organization of the recorded screen content, extracting it into concise summaries, and saving them in Markdown file format on the user's local device. To balance efficiency and storage pressure, these recorded files exist only temporarily, and according to official statements, the system will automatically delete them physically six hours after generating the memory.

image.png

Currently, this cutting-edge feature is available for preview to ChatGPT Pro subscribers on the macOS platform. Eligible users just need to go to the "Personalization" option in settings, check the "Memory" and "Chronicle" features in sequence, and grant the system the corresponding screen recording and accessibility permissions to start experiencing it. However, due to factors such as local privacy laws, this feature is temporarily not available to users in the EU, UK, and Switzerland regions.

Although Chronicle greatly enhances collaboration efficiency, its potential risks should not be ignored. The official released a special reminder: due to the high-frequency data processing involved, it will accelerate the consumption of the user's usage quota. More importantly, there are security challenges—since the memory content is stored locally in an unencrypted form, and the displayed web content may contain malicious instructions, users need to be vigilant against potential prompt injection attacks. This new feature, which balances convenience and complex security boundaries, undoubtedly marks an important step for AI moving from "instruction response" to "environmental perception."