Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MemGovern: Enhancing Code Agents through Learning from Governed Human Experiences

Qihao Wang, Ziming Cheng, Shuo Zhang, Fan Liu, Rui Xu, Heng Lian, Kunyi Wang, Xiaoming Yu, Jianghao Yin, Sen Hu, Yue Hu, Shaolei Zhang, Yanbing Liu, Ronghao Chen, Huacan Wang

TL;DR

MemGovern addresses the open-world challenge in autonomous software engineering by transforming cross-repository GitHub repair history into a structured, agent-friendly experiential memory. It introduces an Experience Governance pipeline (selection, standardization, quality control) and a dual-primitive, agentic memory search (searching and browsing) to enable progressive, analogical repair reasoning. The framework yields 135K governed experience cards and achieves an average improvement of 4.65% in SWE-bench Verified bug-resolution rates across multiple LLM backbones, validating cross-repository open-world learning for code agents. Key contributions include the decoupled Index/Resolution layers, a check-list quality loop, and the agentic search workflow that robustly filters retrieval noise. The work demonstrates practical integration as a plug-in module and highlights the potential to scale agent performance by leveraging curated human repair knowledge from GitHub.

Abstract

While autonomous software engineering (SWE) agents are reshaping programming paradigms, they currently suffer from a "closed-world" limitation: they attempt to fix bugs from scratch or solely using local context, ignoring the immense historical human experience available on platforms like GitHub. Accessing this open-world experience is hindered by the unstructured and fragmented nature of real-world issue-tracking data. In this paper, we introduce MemGovern, a framework designed to govern and transform raw GitHub data into actionable experiential memory for agents. MemGovern employs experience governance to convert human experience into agent-friendly experience cards and introduces an agentic experience search strategy that enables logic-driven retrieval of human expertise. By producing 135K governed experience cards, MemGovern achieves a significant performance boost, improving resolution rates on the SWE-bench Verified by 4.65%. As a plug-in approach, MemGovern provides a solution for agent-friendly memory infrastructure.

MemGovern: Enhancing Code Agents through Learning from Governed Human Experiences

TL;DR

MemGovern addresses the open-world challenge in autonomous software engineering by transforming cross-repository GitHub repair history into a structured, agent-friendly experiential memory. It introduces an Experience Governance pipeline (selection, standardization, quality control) and a dual-primitive, agentic memory search (searching and browsing) to enable progressive, analogical repair reasoning. The framework yields 135K governed experience cards and achieves an average improvement of 4.65% in SWE-bench Verified bug-resolution rates across multiple LLM backbones, validating cross-repository open-world learning for code agents. Key contributions include the decoupled Index/Resolution layers, a check-list quality loop, and the agentic search workflow that robustly filters retrieval noise. The work demonstrates practical integration as a plug-in module and highlights the potential to scale agent performance by leveraging curated human repair knowledge from GitHub.

Abstract

While autonomous software engineering (SWE) agents are reshaping programming paradigms, they currently suffer from a "closed-world" limitation: they attempt to fix bugs from scratch or solely using local context, ignoring the immense historical human experience available on platforms like GitHub. Accessing this open-world experience is hindered by the unstructured and fragmented nature of real-world issue-tracking data. In this paper, we introduce MemGovern, a framework designed to govern and transform raw GitHub data into actionable experiential memory for agents. MemGovern employs experience governance to convert human experience into agent-friendly experience cards and introduces an agentic experience search strategy that enables logic-driven retrieval of human expertise. By producing 135K governed experience cards, MemGovern achieves a significant performance boost, improving resolution rates on the SWE-bench Verified by 4.65%. As a plug-in approach, MemGovern provides a solution for agent-friendly memory infrastructure.
Paper Structure (28 sections, 3 equations, 11 figures, 2 tables)

This paper contains 28 sections, 3 equations, 11 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. MemGovern
  4. Experience Governance
  5. Hierarchical Experience Selection
  6. Experience Standardization
  7. Checklist-Based Quality Control
  8. Experiential Memory Search
  9. Dual-Primitive Interface
  10. Progressive Agentic Search
  11. Selective Browsing and Analogical Transfer.
  12. Experiments
  13. Experimental Setup
  14. Configuration
  15. Main Results
  16. ...and 13 more sections

Figures (11)

  • Figure 1: Performance comparison of SWE-Agent(the base framework) and MemGovern across different LLM backbones on SWE bench Verified.
  • Figure 2: Comparison of MemGovern with existing methods. MemGovern learns from human experience by governing raw data into agent-friendly memories.
  • Figure 3: Architecture of MemGovern. MemGovern selects raw human experiences from GitHub and standardizes them into experience cards, enabling agents to utilize them through agentic experience search.
  • Figure 4: Results of MemGovern under various experiential memory size and MemGovern under various retrieval sizes.
  • Figure 5: Comparison of using raw unprocessed experience and MemGovern's experience.
  • ...and 6 more figures