How Claude Code Works

I've been wrestling with this exact state management problem in a LangGraph implementation recently. Relying on a flat CLAUDE.md file feels like a pragmatic hack to avoid shipping a local database, but the token economics seem rough. If the agent has to re-ingest that entire context file on every turn of the gather-act-verify loop, you are going to burn through input tokens incredibly fast, especially with Opus. I eventually moved my context store to Postgres just to keep the margins viable.