Codex Can Now Search the History of Long Sessions

Codex changes how it handles long context

What stands out here is not the word compact itself, but a change in the memory mechanism: Codex has gained an experimental mode with structured notes and search across older session history. The ChatGPT/Codex changelog and OpenAI configuration reference say the mode preserves accumulated details between context windows and adds the new_context tool. Instead of producing yet another summary of the entire conversation, the agent can start a fresh window and retrieve earlier messages or tool outputs when needed.

In a user message, the mode is enabled with the command codex features enable context_management.experimental_mode. The documented configuration option sets features.context_management.experimental_mode to true. As of September 5, 2026, the feature is off by default and available only to eligible ChatGPT Plus, Pro, and Pro Lite sessions running on the Codex backend, not sessions using an API key or third-party providers.

The logic is sound. Notes retain stable facts, while search retrieves details that were not included in them: an earlier requirement, a planning decision, or a tool conclusion. For autonomous coding, this is more practical than the compress, forget part of it, and continue cycle, especially when work spans multiple context windows.

One user launched an overnight run to test the mode on a long context. Another specifically noted that the output became noticeably easier to read: research plans and complex summaries can now be reviewed without extra skills. There are no official quality figures for experimental mode yet, though, so this remains a strong design idea and an early UX signal rather than a benchmark-proven improvement.

Long agent tasks gain more durable memory

The core change is simple: the agent no longer has to fit all useful history into a single summary that is continuously rewritten. That reduces the risk of losing a decision, constraint, or tool result simply because it did not survive the next compression step. Long refactors, research tasks, and debugging loops with changing plans should benefit most.

I would first look at retrieval accuracy and the quality of the notes themselves. If the agent stores noise or retrieves a formally similar but outdated fragment, long memory can quickly become long confusion. A separate question is how predictably the model decides when to call new_context and which facts should be carried forward.

This is not a magical infinite context window. Rather, Codex has gained a more engineered way to live with a finite window without pretending that one summary is always enough. The main unresolved question is no longer whether the agent can remember, but how accurately it understands what is worth remembering.

We previously covered how code-map interfaces improve context injection and navigation for AI-assisted development. Those same context-selection challenges sit at the center of Codex’s experimental mode for long-running agents.