One night I hit the token limit with Codex and realized most of the cost was coming from context reloading, not actual work. So I started experimenting with a small context engine around it: – persistent memory – context planning – failure tracking – t…