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.
More from coding & agent
- Agent browser silent failures: tool returns success but side effect never lands — Downtown_Item_9996 · 2026-08-11
- Open-Source SkillNet: Search, Install, and Evaluate Reusable AI Agent Skills — tom_doerr · 2026-08-11
- Building a Ground-Up AI E-commerce Creative System: Mockups, Evals, and Agents — aishashok14 · 2026-08-11
- Hands-on with Claude Code: Auto-generating Follow-up Task Cards — dotey · 2026-08-11
- Reconstructing 3D Scenes as Code: Agent Workflow for Real-time Editing — joemeno · 2026-08-11
- BlackHat Warning: 171 Ways for AI Agents to Escape Sandbox via Communications — Ghost_Pilot_MD · 2026-08-11