Model completely loses context thread and repeats previous response

What version of the Codex App are you using (From “About Codex” dialog)?

Latest version

What subscription do you have?

Plus, Business

What platform is your computer?

_No response_

What issue are you seeing?

During a session using the "GPT 5.4 extra high" model, with a context window of ~350k tokens, the AI completely lost track of the conversation.

I sent a new prompt asking a specific question about directory structures. Instead of answering the new prompt, the model either failed to register the new message entirely or simply repeated a previous response about an older, already-resolved topic (terminal rendering and network checks). It essentially lost the thread of the conversation, (and this is not the first time that such a thing is happening...)

When I explicitly pointed out that it was answering the wrong thing, it acknowledged the error, literally stated "I lost the thread," and then finally answered the correct prompt.

I don't know if this is a bug with the Codex application's context manager (feeding the wrong/old message history to the model) or if it is an inherent issue with the model itself losing track of the context at ~350k tokens.

What steps can reproduce the bug?

1. Run a long session using the "GPT 5.4 extra high" model.
2. Accumulate around ~350k tokens of context history.
3. Send a new prompt that switches to a different topic.
4. Observe: The AI ignores the new prompt and repeats its previous response or replies or generate content regarding to an older message.
5. Correct the AI, and observe it successfully recover and apologize for "losing the thread."

What is the expected behavior?

_No response_

Additional information

_No response_