AgentCL

AgentCL: if the task stream is not controlled, agent memory gains prove nothing

The framework pairs compositional task streams against naive ones and reads the gap. Six existing benchmarks are set aside, and the memory designs tested sometimes made performance worse.