Recursive Experiential-Working Memory Evolution for Long-Horizon Agent Harnesses