SWE-Exp: Experience-Driven Software Issue Resolution

Recent advances in large language model (LLM) agents have substantially improved software issue resolution by enabling models to interact with codebases, reason over issue contexts, and iteratively produce patches. Yet existing memory-enhanced software engineering agents mainly organize prior repair knowledge as repository-specific memory, leaving open whether experience learned from one codebase can help resolve issues in another codebase. This paper studies cross-repository experience transfer for repository-level issue resolution. We propose SWE-Exp, an experience-driven framework that distills concise, actionable issue-type knowledge from past repair trajectories and reuses it across repositories. SWE-Exp builds a multi-faceted experience bank from both successful and failed attempts, capturing reusable knowledge ranging from high-level problem understanding to concrete code modification strategies. It further introduces a dual-agent design that separates experience-guided planning from execution-level decision making. Experiments on SWE-Bench Verified show that SWE-Exp consistently outperforms prior agent frameworks across three model settings, achieving Pass@1 scores of 42.0% with DeepSeek-V3-0324, 69.0% with DeepSeek-V3.2, and 73.0% with Claude 4 Sonnet. These results show that repair experience can transfer across codebases when it is abstracted beyond repository-specific details.