Model recall degrades as context grows, so a 1M-token window is not 1M tokens of usable working memory.
The fix is a token budget per step, not a bigger model.
Just-in-time loading, external notes and summarisation each reclaim tokens, and each costs something specific.