Memory distillation lifts a 4B model 27 points on AppWorld

Agent Memory Distillation lifted a 4-billion-parameter model's AppWorld score by 27.2 points, per a paper posted to arXiv (2608.07169).

The technique splits an agent's memory into three layers instead of one context dump. Workflow memory holds task strategy. Subtask memory holds intermediate behavioral examples. Function memory holds calling conventions and a list of common errors. GPT-5-mini served as the teacher model that generated all three, then the memory got distilled down for smaller models to use.

That split matters for anyone running agents on hardware that can't fit a frontier model. The teacher only has to generate the memory once. After that, a 4B model gets most of the benefit without needing GPT-5-mini in the loop at inference time.

Workflow and subtask memory get injected before the agent starts a task. Function memory works differently. It only fires after a tool call fails, as a reactive fix rather than a standing instruction.

The numbers back up that structure. Beyond the 27.2-point AppWorld gain, the same setup added 11.2 points on BFCL V3 and 3.4 on ToolSandbox. Subtask memory, the middle layer, contributed the most of the three, and 4B models saw the biggest lift of any size tested.

The paper doesn't say what generating the memory costs, or whether the gains hold with a different teacher model.

Concrete behavioral examples beat abstract task strategy for small models here. If that pattern holds under ablation on other benchmarks, giving an agent examples looks like a better default than giving it a plan.