memory layers vs plain long context: setup advice

[quiethouse]

Thanks, that helps. The pattern around KV cache pressure has been pretty similar for me. The biggest change for me was simplifying the stack instead of adding one more optimization layer.