Agent Memory Distillation: Empowering Small LLM Agents with Hierarchical Teacher Memory