MemPlan: A Memory-Conditioned PDDL Planning Framework for Partially Observable Text Environments
Summary:MemPlan is a novel PDDL planning framework for partially observable interactive text environments, which leverages memory-conditioning mechanisms to optimize planning efficiency. The framework dynamically adjusts the costs and constraints of candidate hypotheses through positive and negative memory modules, effectively reducing online token consumption and improving task completion rates. Experiments show that MemPlan outperforms language-only and traditional PDDL-based baselines across multiple
MemPlan: A Memory-Conditioned PDDL Planning Framework for Partially Observable Text Environments
1. Core Breakthroughs
MemPlan is a novel PDDL (Planning Domain Definition Language) planning framework designed for partially observable interactive text environments. Its core innovations include:
-
Memory-Conditioned Mechanism: The framework uses positive and negative memory modules to dynamically adjust the costs and constraints of candidate hypotheses during the planning process. The positive memory module assigns relevance-based action costs to candidate hypotheses for missing objects, locations, or conditions, guiding the planner toward more plausible checks. The negative memory module records execution-refuted assumptions and compiles them into exclusion constraints, preventing repeated attempts at candidates already contradicted by feedback.
-
Unified Language-Symbol Interface: A unified schema-conditioned interface, trained offline with supervised fine-tuning, connects text environments to the planning loop, enabling the transformation from observation parsing to fact extraction and then to symbolic-action-to-command grounding.
2. Technical Mechanism Analysis
MemPlan’s architecture addresses the challenges of using LLMs as long-horizon planners, which are costly and prone to hallucinated or invalid actions. Its main technical mechanisms include:
-
Dynamic Problem Reconstruction: At each step, MemPlan reconstructs the PDDL problem based on parsed observations, candidate hypotheses, and interaction memory, rather than relying on a fixed PDDL problem description. This approach enhances the flexibility and adaptability of the planning process.
-
Memory-Guided Planning Optimization: The collaborative work of positive and negative memory modules enables the planner to make more effective use of historical information, reducing unnecessary computations and repeated attempts, thereby lowering online token consumption.
3. Engineering Trade-offs and Performance
MemPlan performs well across multiple benchmarks, including TextWorld, ALFWorld, and Robotouille. Compared to language-only and traditional PDDL-based baselines, MemPlan significantly improves task completion rates and step efficiency. Additionally, compared to LLM planning baselines like ReAct, MemPlan substantially reduces online token costs.
However, MemPlan’s limitations include its high demand for offline training data and potential challenges in handling highly dynamic or unpredictable environments.
4. Developer Deployment Recommendations
For developers, MemPlan provides an effective method for combining LLMs with symbolic planning. Here are some deployment recommendations:
-
Data Preparation: Ensure sufficient offline training data to train the unified language-symbol interface.
-
Memory Management: Since MemPlan relies on memory modules, developers need to optimize memory usage to adapt to different application scenarios.
-
Environment Adaptation: Adjust the PDDL domain definition and observation parsing logic according to the specific application environment to fully leverage MemPlan’s advantages.
MemPlan’s code and data are available on GitHub at: https://github.com/tewiSong/MemPlan.
— END —Source: ArXiv NLP/LLM (cs.CL) (2026-10-24)
Tags: #MemPlan #PDDL #LLMs & Foundation Models #Intelligent Agents #Text Environments
Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.
Community Comments