Researchers propose a new framework called SkillCommit to improve the performance of large language models (LLMs) without updating their parameters. The framework transforms historical experience into reusable procedural knowledge by abstracting related skills into higher-level skills. This is done through embedding-based retrieval and cross-instance replay, which check whether skills transfer across cases and share a common underlying mechanism. Experiments show that SkillCommit improves agent performance across diverse domains and enables cross-model experience transfer. --- Why it matters: This matters to researchers in AI because it addresses the issue of consolidating experience based on semantic similarity or LLM judgments, which can degrade performance by merging superficially related but behaviorally incompatible strategies. Source: https://arxiv.org/abs/2608.15165