OpenAI Codex Chronicle: Your Screen History Becomes Its Long-Term Memory

OpenAI’s Codex now watches your screen. Chronicle runs in the background, periodically captures screenshots, summarizes them using a sandboxed Codex session, and stores the result as local markdown files. The next time you open Codex, it already knows what you’ve been working on.

Context

The persistent friction in AI-assisted coding isn’t model capability — it’s cold-start context. Every new session requires the user to reconstruct their working environment for the agent: here’s the codebase, here’s the task, here’s the constraint. That explanation overhead compounds across dozens of daily interactions, and it’s proportionally more expensive for complex, long-running projects where the context needed to be useful is large.

The memory layer has emerged as the strategic differentiator in coding agent infrastructure. An agent that maintains persistent awareness of what the user has been working on doesn’t need to be briefed; it can contribute immediately. Chronicle is OpenAI’s architecture for that problem, and the approach it uses — passive local summarization rather than cloud-synchronized memory — defines both its capability ceiling and its privacy boundaries.

The Detail

Chronicle operates as a background process that captures screen recordings at configurable intervals. Sandboxed Codex sessions process those recordings into summaries stored as unencrypted markdown files in a local directory ($CODEX_HOME/memories_extensions/chronicle/). The memories become available context for subsequent Codex sessions, reducing the amount of re-explanation required each time the user opens a new session. Recordings older than six hours are automatically deleted; the memories they generated persist.

OpenAI’s own documentation lists the tradeoffs directly. The unencrypted local storage is a meaningful attack surface on shared machines or in environments where local file access isn’t strictly controlled. The feature “raises the risk of prompt injection attacks” — malicious instructions embedded in a displayed website can, in theory, be captured in a screenshot, summarized into the memory store, and influence subsequent Codex behavior. Chronicle also consumes rate limits at above-average speed, since the background summarization competes with active usage budgets.

Availability is currently restricted to ChatGPT Pro subscribers on macOS, with the EU, UK, and Switzerland excluded — a geographic restriction that reflects GDPR compliance complexity around passive screen monitoring rather than a permanent boundary. The feature is opt-in, requiring macOS screen recording and accessibility permissions before activation.

The positioning Chronicle occupies in OpenAI’s agentic product strategy is notable: it’s a memory architecture that doesn’t require cloud synchronization. The privacy argument for local storage is real — screen captures of a developer’s active projects contain sensitive information that most users wouldn’t want transmitted to a third-party server. The six-hour deletion window on raw captures limits long-term exposure while allowing the summarized memories to accumulate. The tradeoff is that context depth is bounded: captures from last week don’t inform today’s session, only the summarized memories they generated do.

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.