When to Start Fresh: Managing AI Coding Sessions

Why context degrades

Every message in a conversation is part of Claude Code's working memory. Early in a session, that memory is clean — all of it is relevant to what you're doing right now. But as a session grows, it accumulates dead weight: earlier approaches you abandoned, half-finished ideas you explored and discarded, tangents you pursued and closed.

Claude Code can't distinguish between 'this was relevant before' and 'this is relevant now.' It weighs everything together. As the irrelevant content piles up, it dilutes the signal from the current task. The AI starts mixing up what it said before with what you're asking now, and contradicting itself.

This isn't a bug. It's how language models work. The fix is knowing when to clear the slate.