Agent Memory Distillation: Empowering Small LLM Agents with Hierarchical Teacher Memory
The episode discusses a paper from KAIST and DeepAuto.ai on Agent Memory Distillation, which transfers a large teacher model's memory to small student models via three hierarchical memory types: workflow, subtask, and function. This training-free method boosts small models' accuracy on benchmarks like AppWorld, sometimes surpassing the teacher.