Core Insights - The article discusses the limitations of current AI in bug fixing and introduces the MemGovern framework, which aims to enhance AI's ability to learn from human experiences in software engineering [2][3][28]. Group 1: AI Limitations and Challenges - Current AI systems struggle to effectively utilize the vast amount of historical data available on platforms like GitHub due to their "closed world" cognitive limitations [3][4]. - Human engineers often rely on community knowledge to solve complex issues, but AI's direct access to unstructured data from GitHub is hindered by noise and ambiguity [4][8]. Group 2: MemGovern Framework - MemGovern introduces an "Experience Refinement Mechanism" that transforms raw GitHub data into structured, AI-friendly "experience cards" [9][10]. - The framework employs a hierarchical selection process to filter high-quality repositories and cleanses data to retain only complete repair records [9][10]. Group 3: Experience Cards - Experience cards consist of two layers: an Index Layer for efficient symptom-based retrieval and a Resolution Layer that encapsulates root cause analysis, fix strategies, patch summaries, and verification methods [10][12]. - The structured design of experience cards enhances the usability of knowledge and allows for better retrieval of repair logic [10][12]. Group 4: Search Mechanism - MemGovern utilizes a "Search-then-Browse" approach, allowing AI to first search for relevant cases based on symptoms and then browse detailed solutions, mimicking human search behavior [12][13]. - This method enables AI to understand repair logic more deeply and filter out irrelevant information [12][13]. Group 5: Experimental Results - MemGovern has shown significant improvements in bug resolution rates across various models, with Claude-4-Sonnet+MemGovern achieving a 69.8% resolution rate, a 3.2% increase over the baseline [15][16]. - GPT-4o+MemGovern's resolution rate increased from 23.2% to 32.6%, marking a 9.4% improvement [16]. Group 6: Broader Implications - The MemGovern framework not only enhances performance metrics but also provides a clear pathway for AI agents to effectively utilize vast amounts of unstructured human debugging experience [28]. - The methodology has potential applications beyond coding, such as in legal consulting and medical diagnosis, where historical case knowledge is crucial [28].