Codex can forget things as planned, rather than losing the project

The most dangerous situation for an AI agent isn’t a lack of context in and of itself, but rather a loss of coherence in the middle of a large task. In the proposed patch, Codex compact becomes not an emergency compression but a controlled pause: the agent explains the reason in advance, saves its plan, and only then continues working.

Context Becomes a Checkpoint

Before a major phase, the agent checks whether the remaining time window is sufficient. If not, it saves a checkpoint: what has already been done, what is currently being worked on, and what should happen after the context is compressed.

After compaction, the model reads this checkpoint, re-checks the repository’s state, and returns to the plan. This is an important change in logic: the source of truth is not only the model’s memory but also the verified state of the project.

Why This Is More Practical Than Regular Memory

The agent gains the ability to work through long sequences of actions without pretending that the context is infinite. It doesn’t have to keep the entire history in mind—it’s enough to save its immediate intentions and check against the code after restarting.

But there’s a downside: the more decisions the agent makes on its own, the more important it is to require an explanation and an explicit record of its plan. Otherwise, compact will simply hide an error behind a neat retelling of past actions.

For now, this is just an experiment

The author explicitly states that no conclusions have been drawn following testing yet. To test it, you need to patch Codex; you can try out the implementation in codex-luna-kit. The source provides no data on the approach’s stability, comparisons with the standard compact command, or its performance in real-world projects—so it’s too early to call this a reliable solution.

But the idea itself is sound: an agent doesn’t need “eternal memory,” but rather a procedure for safe forgetting. Would you trust an agent to independently choose when to compact if it’s required to leave a checkpoint and re-verify the repository?

Source: All Articles in a Row / Artificial Intelligence / Habr