The Problem You finish a session with Claude Code at 11pm. Three files changed, two design decisions made, one bug discovered but unresolved. You start fresh the next morning. The agent has no memory of any of it. You spend the first ten minutes of the new session doing what I call "cold-start theater": re-reading the changed files, re-explaining what you decided yesterday, re-discovering the bug you already debugged. Multiply that by every session. That's the tax you pay for stateless agents. Why Conversation History Doesn't Solve This The obvious answer is "just save the conversation." It doesn't work for three reasons: Conversation history is enormous. A real working session has thousands of messages, tool calls, and outputs. Loading it consumes most of your context window before you do anything useful. It's mostly noise. 90% of what happened in the previous session was the agent reasoning out loud. The next agent doesn't need that. It needs the conclusions. It doesn't compose across agents.…