Intel / article

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

Source: arxiv.org
Author: Jiahao Zhang, Yifan Zhang, Yu Huang
Date: Published Aug 10, 2026

Context

PMCoder links a coding 's plan to its memory in both directions. The current repair phase determines which past observations it retrieves, while patterns in that history help it detect when it is stuck and revise the plan. This targets long repair attempts in which an agent repeats failed edits or keeps working from stale evidence. When available, executed issue-reproduction results also inform verification progress. Research abstract

On , the authors report an average of 25 additional resolved cases, a five-percentage-point gain over a baseline using a matched agent setup. Their component tests found that combining planning and memory worked better than either alone. The result supports coordinating these capabilities rather than merely adding a larger store of past observations. Reported evaluation

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
SWE-bench — The standard benchmark for AI coding agents: real GitHub issues from real repositories, scored by whether the agent's patch passes the project's own tests.
eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.

Read the source arxiv.org