RPMS: Enhancing LLM-Based Embodied Planning through Rule-Augmented Memory Synergy

2026-03-18 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Advanced, quick

Summary

RPMS (Rule-Augmented Memory Synergy) is a novel architecture designed to improve Large Language Model (LLM) agents' performance in closed-world embodied environments where actions require strict precondition satisfaction and feedback is sparse. It addresses two primary failure modes: invalid action generation and state drift, which often amplify each other. RPMS integrates structured rule retrieval to enforce action feasibility, employs a lightweight belief state to gate memory applicability, and uses a rules-first arbitration mechanism to resolve conflicts between these sources. Evaluated on ALFWorld with 134 unseen tasks, RPMS achieved 59.7% single-trial success with Llama 3.1 8B, a 23.9 percentage point increase over the baseline, with rule retrieval contributing 14.9 percentage points of this gain. It also reached 98.5% success with Claude Sonnet 4.5 (+11.9 pp). The architecture demonstrated transferability to ScienceWorld with GPT-4, yielding an average score of 54.0 compared to 44.9 for the ReAct baseline.

Key takeaway

For AI scientists developing LLM agents for embodied environments, integrating RPMS's conflict-managed architecture can significantly boost task success rates. You should prioritize implementing structured rule retrieval and a belief-state-filtered memory system to mitigate invalid actions and state drift, especially in environments with strict preconditions and sparse feedback. This approach improves reliability and transferability across diverse tasks.

Key insights

RPMS enhances LLM embodied planning by enforcing action feasibility and managing memory through rules-first arbitration.

Principles

• Action preconditions are critical for LLM agent success.
• Episodic memory requires grounding to be beneficial.
• Rules-first arbitration resolves conflicting information sources.

Method

RPMS uses structured rule retrieval for action feasibility, a belief state for memory gating, and rules-first arbitration to manage conflicts, addressing invalid actions and state drift.

In practice

• Implement rule retrieval for action precondition enforcement.
• Filter episodic memory based on current belief state.
• Prioritize explicit rules over memory in conflict resolution.

Topics

• Embodied AI
• LLM Agents
• Planning Architectures
• Rule-Augmented Memory
• Action Feasibility

Best for: AI Scientist, Research Scientist, AI Researcher, AI Engineer, Machine Learning Engineer

Related on AIssential

• Agents Fail to Reject Stale Memories — LLM efficiency, steerability, and agent memory coherence face fundamental limitations requiring novel architectural solutions.
• Experience Compression Spectrum: Unifying Memory, Skills, and Rules in LLM Agents — Experience compression unifies LLM agent memory, skills, and rules, crucial for scalable, long-horizon deployments.
• WorldLines: Benchmarking and Modeling Long-Horizon Stateful Embodied Agents — Long-horizon embodied agents require benchmarks and memory frameworks that handle dynamic, stateful, and partially observable environments.
• From Prompts to Pavement Through Time: Temporal Grounding in Agentic Scene-to-Plan Reasoning — Temporal grounding in AV LLM/LMM ensembles reshapes reasoning but doesn't statistically improve standard correctness metrics, yet shows qualitative benefits.
• From Digital to Physical: Digital Agents as Autonomous Coaches for Physical Intelligence — LLM agents can autonomously engineer embodied policies, outperforming human baselines and self-correcting through iterative environment feedback.
• A robot operating system framework for using large language models in embodied AI — Connecting LLM agents to ROS enables versatile, robust, and scalable embodied AI through natural language.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.