Agent Memory Distillation boosts small LLM agents by up to 27.2% accuracy via hierarchical teacher memory

kaist-ai · hf · 2026-08-11

KAIST AI proposes Agent Memory Distillation (AMD), a training-free framework that transfers structured knowledge from a large teacher agent to small student agents via hierarchical memory. AMD builds three memory types: workflow (task-level strategies), subtask (concrete behavioral examples), and function (per-function calling conventions). Workflow and subtask memories are injected proactively, while function memory is retrieved reactively on errors. On three tool-use benchmarks with GPT-5-mini as teacher, four student models (4B-8B) achieve average accuracy gains of 27.2%p, 11.2%p, and 3.4%p on AppWorld, BFCL V3, and ToolSandbox, outperforming baselines. Subtask memory contributes most; 4B students benefit most.

Original post →

More from coding & agent

coding & agent channel →