Context rot is not the same as running out of room. It sets in well before the hard limit, because a model weights recent tokens more heavily than distant ones. Instructions you set in your first message are technically still in context but assigned low salience, so the model reads them and effectively ignores them.

The symptoms are recognizable: a format rule you established early quietly stops being followed, a constraint you set gets violated, or the assistant re-asks something you already answered. Anthropic describes the effect in its own guidance on long Claude Code sessions, noting that performance degrades as context grows because attention is spread across more tokens and older content starts to distract from the current task.

The practical countermeasures are all forms of re-anchoring. Restate core instructions every ten or so exchanges, put critical constraints near the end of your message where they carry more weight, and split genuinely unrelated tasks into separate conversations rather than stacking them into one thread.