Agent Memory Distillation: Empowering Small LLM Agents with Hierarchical Teacher Memory
A training-free memory handoff let 4B-8B agent models gain the most on tool-use tasks when learning from successful GPT-5-mini trajectories.
The paper introduces Agent Memory Distillation, which builds workflow, subtask, and function memories from a larger teacher agent’s successful runs. Workflow and subtask memories are supplied before a task begins, while function memory is retrieved after tool-calling errors. Across AppWorld, BFCL V3, and ToolSandbox, the authors report average accuracy gains of 27.2, 11.2, and 3.4 percentage points. They say subtask memory drove the largest improvements, with 4B students benefiting most. HF Daily Papers' note
The paper introduces Agent Memory Distillation, which builds workflow, subtask, and function memories from a larger teacher agent’s successful runs. Workflow and subtask memories are supplied before a task begins, while function memory is retrieved after tool-calling errors. Across AppWorld, BFCL V3, and ToolSandbox, the authors report average accuracy gains of 27.2, 11.2, and 3.4 percentage points. They say subtask memory drove the largest improvements, with 4B students benefiting most. HF Daily Papers' note
score 4