MemPlan: Memory-Conditioned PDDL Planning Framework for Partially Observable Text Environments Released

Summary:MemPlan is a novel memory-conditioned PDDL planning framework designed to address the high cost and hallucination issues of using Large Language Models (LLMs) as long-horizon planners. By keeping the PDDL domain fixed and reconstructing the PDDL problem at each step from parsed observations, candidate hypotheses, and interaction memory, MemPlan employs positive and negative memory mechanisms to guide the planner toward more plausible actions. Experiments on TextWorld, ALFWorld, and Robotouille e

---

Technical Mechanism Analysis

The core innovation of MemPlan lies in its memory-enhanced PDDL planning framework, which operates through the following mechanisms:

1. Fixed PDDL Domain and Dynamic Problem Reconstruction: MemPlan keeps the PDDL domain fixed and reconstructs the PDDL problem at each step based on parsed observations, candidate hypotheses, and interaction memory. This approach ensures the consistency of planning while adapting to the dynamic nature of partially observable environments.

2. Positive and Negative Memory Mechanisms:

   • Positive Memory: Assigns relevance-based action costs to candidate hypotheses, guiding the planner toward more plausible actions.
   • Negative Memory: Records execution-refuted assumptions and compiles them into exclusion constraints, preventing repeated attempts at candidates already contradicted by feedback.

3. Unified Schema-Conditioned Language-Symbol Interface: Connects text environments to the planning loop through an interface trained offline with supervised fine-tuning, enabling observation-to-fact parsing and symbolic-action-to-command grounding.

Engineering Trade-offs and Performance

MemPlan demonstrates strong performance in multiple benchmarks:

• Task Completion Rate: MemPlan significantly outperforms language-only and PDDL-aware baselines in terms of task completion rate across TextWorld, ALFWorld, and Robotouille environments.
• Efficiency: MemPlan shows superior step efficiency, particularly in complex tasks where its planning paths are more efficient.
• Online Token Cost: Compared to LLM-planning baselines like ReAct, MemPlan substantially reduces online token costs, showcasing its advantages in resource-constrained environments.

However, MemPlan's limitations include its reliance on pre-trained language models and PDDL domains, which may restrict its application in non-textual environments or complex symbolic reasoning tasks.

Developer Adoption and Deployment Recommendations

For developers looking to adopt MemPlan, here are some recommendations:

1. Environment Adaptation: Ensure the target environment meets MemPlan's requirements for textual interaction and partial observability.
2. Domain Knowledge Integration: Integrate domain knowledge into the PDDL domain definition to enhance planning efficiency and accuracy.
3. Memory Management: Configure positive and negative memory mechanisms appropriately to balance planning precision and computational cost.
4. Interface Optimization: Optimize the language-symbol interface according to specific application scenarios to improve overall system performance.

The open-source code and data for MemPlan are available on GitHub, providing a solid foundation for experimentation and extension by developers.

---

Source: ArXiv NLP/LLM (cs.CL) (2026-10-23)

— END —

Tags: #MemPlan Follow #PDDL Follow #LLMs & Foundation Models Follow #Intelligent Agent Planning Follow #Partially Observable Environments Follow