Agent Memory Distillation lifts a 4B model 27.2 points on AppWorld by transferring three tiers of teacher memory

arXiv·medium signalXBlueskyLinkedInCopy link

Submitted 2026-08-07 (arXiv:2608.07169), AMD distills a large teacher's experience into three memory layers for small 4B–8B agents: workflow memory (task strategy), subtask memory (intermediate behavioral examples), and function memory (calling conventions and common errors). Workflow and subtask memories are injected proactively while function memory fires reactively only on a tool-calling error. With GPT-5-mini as teacher, gains were 27.2 points on AppWorld, 11.2 on BFCL V3, and 3.4 on ToolSandbox — subtask memory contributed most, and 4B models benefited most, which is the practical read: the smaller your local agent, the more this buys you.