Skill-Synthesizer: Query-Aware Skill Synthesis for Scientific Reasoning

Maximizing the value of an agent's accumulated experience requires more than retrieving the most similar skill: it requires selectively combining and adapting knowledge based on the demands of the current problem. This is especially true in scientific reasoning, where solving a problem often requires integrating theoretical principles and procedural knowledge from multiple sources. Existing skill-based memory systems fall short of this: they retrieve whole skills and inject them directly into the agent's context, without reasoning over their content or relationships. This is suboptimal when relevant knowledge is scattered across multiple skills, or when retrieved skills contain information only partially relevant to the current task. We present \textbf{\skillsyn}, which addresses these limitations by retrieving relevant passages from a skill library and composing them into a unified, query-specific skill. A dedicated synthesis agent reasons over how retrieved passages complement each other, assembles them into an executable procedure, and verifies intermediate steps through code execution, a highly valuable capability for the derivation-heavy problems that characterize scientific workflows. Evaluated on two challenging physics benchmarks, \skillsyn consistently outperforms all retrieval baselines in both accuracy and reasoning soundness, with especially strong gains on a graduate-level benchmark where no other retrieval method improves over the memoryless baseline.