MemPlan is a memory-augmented PDDL planning framework designed for partially observable interactive textual environments, first made public on October 24, 2026, as arXiv:2610.30379v1. This work does not employ large language models (LLMs) directly as long-range planners; instead, it positions them as a “restricted bridge” that connects natural-language observations with symbolic planning.
The core of the framework is to keep the PDDL domain definition fixed, while dynamically reconstructing the PDDL problem instance at each step based on parsed observations, candidate hypotheses, and interactive memory. Specifically, forward‑graph memory assigns action costs to candidate hypotheses for missing objects, locations, or conditions according to their relevance, guiding the planner to prioritize more plausible hypotheses; negative memory, on the other hand, records hypotheses that have been refuted by execution feedback and encodes them as exclusion constraints, thereby preventing redundant attempts at options that have already been ruled out.
MemPlan employs a unified modal-conditioned language–symbol interface, which is fine-tuned via offline supervised learning to enable observation-to-fact parsing and symbolic-action-to-command grounding. Experiments on three benchmarks—TextWorld, ALFWorld, and Robotouille—demonstrate its effectiveness: compared with pure language approaches and PDDL-aware baselines, MemPlan improves task success rates and step efficiency; compared with LLM-planning baselines such as ReAct, it significantly reduces online token consumption. Moreover, the same fine-tuning interface can be directly reused across different task types in TextWorld without requiring additional task-specific fine-tuning.
The authors of the paper have made all code and data publicly available, hosted on GitHub (https://github.com/tewiSong/MemPlan）。需注意：所有结论均基于论文所载实验与描述，未引入外部评估或第三方复现结果；发布时间2026-10-24为arXiv预印本发布日期，非期刊或会议正式录用时间。