Researchers have introduced Agent Memory Distillation (AMD), a training-free framework designed to enhance the performance of small language model (SLM) agents by transferring structured knowledge from a large teacher model. Using GPT-5-mini as the teacher, AMD extracts three distinct hierarchical memory types from successful agent trajectories: Workflow memory, which captures high-level task strategies; Subtask memory, which provides intermediate-level behavioral examples; and Function memory, which details specific tool-calling conventions and common error patterns. The framework employs a dual-injection strategy where Workflow and Subtask memories are proactively provided at the beginning of a task, while Function memory is reactively retrieved following tool-calling failures. Experimental evaluations on three benchmarks—AppWorld, BFCL V3, and ToolSandbox—demonstrate that 4B-8B parameter student models achieve significant average accuracy improvements of 27.2%p, 11.2%p, and 3.4%p, respectively. The study reveals that Subtask memory provides the most substantial performance boosts and that the benefits of AMD are most pronounced in smaller 4B-sized models. Furthermore, the effectiveness of the distillation process is contingent upon both the capability of the teacher model and the inherent compatibility of the student model.