Clearing

Clearing throws away the current session's conversation so the next request starts from an empty context window. Only the standing pieces survive: the system prompt, your AGENTS.md file, and whatever the memory system loads on startup. Everything else, including every file the agent read and every decision it made along the way, is gone. In Claude Code the command is /clear; other harnesses call it a new chat or a fresh thread.

Clearing feels like losing work, so people avoid it and push a bloated session onward instead. That is usually the wrong trade. Nothing you cleared was saved anywhere except in the window, so if a fact matters it should already live in a file, a commit, or a note. What a long session mostly holds is stale tool output and abandoned dead ends, and dragging those along is what pushes the agent out of the smart zone.

Clear at task boundaries, not mid-task. Finish the piece of work, commit it, write down anything the next session needs, then clear and start the next piece with a clean window. If the task is too big to finish before quality drops, that is a cue to write a handoff rather than to push through. Prefer clearing over compaction whenever the summary would need to carry more than a couple of sentences.