To provide context, I utilize the memory-bank pattern with GitHub Copilot Agent, but I believe I'm wasting a significant number of tokens.