EXPEREPAIR: Dual-Memory Enhanced LLM-based Repository-Level Program Repair

2025-06-12

AI Summary

Existing repository-level LLM-based program repair methods suffer from two key limitations: (1) isolated problem processing without leveraging historical repair experience, and (2) reliance on static prompts, resulting in poor generalization. To address these, we propose the first human-inspired dual-memory mechanism: an episodic memory that stores concrete repair instances and a semantic memory that distills generalized repair patterns—dynamically fused to generate context-aware prompts. Integrated with retrieval augmentation, memory encoding, and dynamic prompt synthesis, our approach achieves 49.3% pass@1 on SWE-bench Lite using Claude 3.7 Sonnet, significantly outperforming all open-source state-of-the-art methods. Our core contributions are: (1) an evolvable, memory-driven program repair paradigm; (2) a code-specific semantic-episodic collaborative memory architecture; and (3) a fine-tuning-free framework for dynamic, context-sensitive prompt generation.

Technology Category

• Natural Language Processing: Code Generation / Program Synthesis from Natural Language
• Planning, Routing, and Scheduling: Replanning and Plan Repair
• Machine Learning: Large Multimodal Models (LMMs)

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

Automatically repairing software issues remains a fundamental challenge at the intersection of software engineering and AI. Although recent advancements in Large Language Models (LLMs) have demonstrated potential for repository-level repair tasks, current methodologies exhibit two notable limitations: (1) they often address issues in isolation, neglecting to incorporate insights from previously resolved issues, and (2) they rely on static and rigid prompting strategies, which constrain their ability to generalize across diverse and evolving issue scenarios. Inspired by the dual memory systems of human cognition, where episodic and semantic memories work synergistically to support human reasoning and decision-making, we propose ExpeRepair, a novel LLM-based approach that continuously learns from historical repair experiences through dual-channel knowledge accumulation. ExpeRepair organizes historical repair experiences into two complementary memories: an episodic memory that stores concrete repair demonstrations, and a semantic memory that encodes abstract reflective insights. At inference time, ExpeRepair activates both memory systems by retrieving relevant demonstrations from episodic memory and recalling high-level repair insights from semantic memory. It further enhances adaptability through dynamic prompt composition, synergistically integrating both memory types to replace static prompts with context-aware, experience-driven prompts. Experiments on the SWE-bench Lite benchmark demonstrate that ExpeRepair achieves a pass@1 score of 49.3% with Claude 3.7 Sonnet, outperforming all state-of-the-art open-source methods.

Problem

• Addresses limitations in LLM-based repository-level software repair
• Enhances repair accuracy by leveraging dual-memory learning from historical fixes
• Improves adaptability with dynamic prompts over static strategies

Innovation

• Dual-memory system for continuous learning
• Dynamic prompt composition for adaptability
• Repository-level repair with contextual insights