This ties directly to statelessness. Because AI is stateless, these platforms feed the entire conversation back into the model on every message. But there's a limit to how much it can actually hold. That limit is the context window limit.
It's measured in tokens, which are roughly like a character or word count. If you want to see exactly how many tokens a word or paragraph is, OpenAI has a free "Tokenizer" tool you can look up. Once that window fills up, quality drops, and research shows it doesn't drop in a straight line, it gets worse exponentially, so it goes bad fast. That long conversation where AI started making more mistakes? This is exactly why.
Two things you can do about it:
- 1. Handoff document. For a bigger project in Claude Code or Cowork where you don't want to lose everything, have AI write a handoff file for the next chat. The prompt I use: "Create a handoff document that would allow a brand-new chat with zero context to pick up exactly where we left off." It gets you 90 to 95% of the way there, which is plenty.
- 2. Canary in the coal mine. Tell the model to start every message with your first name. As long as it does, the window is still fresh and it's holding your instructions. The moment it drops your name, that's your signal it's starting to forget, so it's time for a fresh chat or a handoff doc.