Can small agents learn from larger agents' memories?

The paper starts from a gap: memory has helped large proprietary agents, but its potential "remains largely unexplored for small language models, which struggle to generate sufficient successful trajectories on their own." The obvious fix is to borrow a stronger agent's trajectories. The authors report that this fails on its own terms: naive memory transfer "yields only marginal improvements due to the capability gap between teacher and student agents," and the gap "cannot be bridged by simply providing high-quality teacher memories." What they propose instead is Agent Memory Distillation (AMD), a training-free framework in which the teacher's successful trajectories are restructured into hierarchical memory before the student sees them. With GPT-5-mini as teacher and four students of 4B to 8B parameters, average accuracy gains are 27.2, 11.2 and 3.4 percentage points on AppWorld, BFCL V3 and ToolSandbox, and AMD beats every memory-based baseline. Some students approach or surpass teacher-level performance.

The mechanism is a split by granularity and by timing. Workflow memory encodes task-level strategies, subtask memory gives "concrete behavioral examples at an intermediate granularity," and function memory records per-function calling conventions and common pitfalls. Workflow and subtask memories are injected proactively at the start of each task. Function memory is retrieved reactively, only when a tool call errors. The discussion adds that each memory type "contributes distinctly" in ablations and that effective distillation "requires careful alignment between memory complexity and student capacity." The implied reading is that a memory is not a portable object. Whether it helps depends on what the consuming model can absorb, so the teacher's raw experience has to be re-cut to the student's level.

This sits against the library's granularity notes in two ways. Does agent memory work better at one level of abstraction? argued that no single abstraction level wins and predicted that a composite system choosing level per task class would beat any single architecture. AMD is a composite of that kind, but it stacks the levels within one domain, tool use, and ties the choice to the reader's capacity rather than to the shape of the domain. Its workflow layer is the same family of idea as Can agents learn reusable sub-task routines from past experience?, placed here beside two finer layers. The case, subtask and tool modules in Can agents learn continuously from experience without updating weights? also give a three-part structure and stay training-free, but there the agent's memory comes from its own experience, while AMD's comes from a different, stronger agent. And Can agents learn better from their failures than successes? shows distillation beating raw trajectories for a single agent, a point the AMD result echoes across a teacher-student pair.

The excerpt leaves a lot open. It gives no per-model results, no size for the naive-transfer baseline beyond "marginal," and no account of how memory complexity is matched to student capacity, which the conclusion names as a requirement. The 3.4-point gain on ToolSandbox is modest next to the AppWorld gain, and nothing here tests a teacher other than GPT-5-mini or tasks other than tool use. The defensible reading is narrow: in these three tool-use benchmarks, restructuring teacher memory by granularity and injection timing worked where handing over the same experience directly did not. Whether the layering or the matching to capacity does most of the work is not something this excerpt can say.