EvoMem: Improving Multi-Agent Planning with Dual-Evolving Memory