PMCoder: Coupling Planning with Episodic Memory in LLM Agents for Software Issue Resolution

Researchers have unveiled PMCoder, a problem-solving agent that integrates a hierarchical phase planner with episodic memory to enhance the efficacy of large language model (LLM) agents in tackling real software problems. This study, published on arXiv (arXiv:2608.06811), confronts the issue of lengthy repair processes—often comprising tens to hundreds of steps—requiring exploration, hypothesis formulation, implementation, and verification. The authors highlight that current repository-level agents typically enhance either planning or memory independently, which leaves extended trajectories susceptible to outdated evidence, repeated unsuccessful edits, and verification based on the agent's assertions rather than actual execution. PMCoder features a bidirectional coupling: the current planning phase influences memory retrieval, while statistics from memory aid in detecting stagnation and replanning. When available, issue-reproduction verdicts support verification progress. This research advances AI-assisted software engineering, especially in automated issue resolution.