PMCoder boosts LLM agent issue resolution by 5pp with episodic memory

LLM agents tackling real software issues often need tens to hundreds of steps spanning exploration, hypothesis, implementation, and verification. Existing repository-level agents typically strengthen planning or memory in isolation, leaving long trajectories vulnerable to stale evidence, repeated failed edits, and verification based on self-reported completion rather than execution evidence. To solve this, the authors introduce PMCoder, an issue-resolution agent that couples a hierarchical phase planner with episodic memory. The coupling is bidirectional: the current plan phase conditions memory retrieval, while memory-derived trajectory statistics inform stuck detection and replanning. When issue-reproduction verdicts are available, PMCoder grounds verification progress in actual execution evidence, reducing false confidence.

On SWE-bench Verified, PMCoder resolves an average of 25 additional cases (+5.0pp) over a harness-matched baseline, with gains persisting even where the reproduction gate never fires. On the larger Verified-500 evaluation, the same positive direction holds across Claude Haiku 4.5, DeepSeek-V4-Flash, and an OpenHands port, yielding at least 14 extra resolved cases (+2.8pp). A separate TerminalWorld evaluation suggests the plan-memory substrate transfers beyond issue reports, hinting at broader agentic applicability. Ablation and trajectory analyses confirm that coupling planning and memory outperforms either component alone, while significantly reducing repeated failed actions, empty-patch exits, and context-window exhaustion — a clear signal that memory-aware planning is the next frontier for autonomous software engineering.