A user put Claude in charge of a fleet of Codex agents, and Claude turned up a history bug. One session ran continuously for two days, leaving behind 110MB of logs and consuming roughly 353 million tokens. Reopen it, though, and Codex only remembers the opening sentence.
None of the actual work was lost. Codex stores the full chat log locally and builds a separate index for fast reads. The issue was that index: it was stuck at an early point in the conversation. The underlying records were all still there — Codex just couldn't see them. The user checked all 439 sessions and found 398 with indexing anomalies.
Soon after, people on Windows Desktop, Remote SSH, and other environments reported similar "false amnesia" symptoms: the underlying logs were fine, but the UI stayed frozen at an old position, with no error message and no self-repair attempt. After deleting the corrupted index, the original poster said Codex re-read a 110MB session in about 20 seconds and recovered all 398 broken sessions.
But later reports complicate the fix. The problem isn't always just the index — the raw log itself can contain duplicate sequence numbers. So blindly deleting the index isn't safe in every case. As of now, the bug is still marked Open, with no linked fix PR in sight.