The loop threw away my memory system, optimizing cache hits on llm models

I set out to solve a cost problem. A long Claude session gets more expensive the more it grows. With Opus, a session sitting around 100k tokens runs me roughly a few dollars; at 200k it's about $10; a