MemCompiler: Compile, Don't Inject -- State-Conditioned Memory for Embodied Agents

2026-05-08

AI Summary

Existing embodied agents typically employ static memory injection mechanisms that struggle to adapt to dynamic environmental states and can even degrade performance. This work proposes MemCompiler, the first state-aware memory compilation framework, which leverages structured state representations to dynamically filter and generate memory-based guidance relevant to the agent’s current execution context through a learnable compiler. MemCompiler innovatively integrates dual channels—explicit textual memory and implicit Soft-Mem—to effectively convey non-verbalizable perceptual information, thereby enhancing decision quality without compromising computational efficiency. Experimental results demonstrate that MemCompiler achieves up to a 129% performance improvement over memory-free baselines across AlfWorld, EmbodiedBench, and ScienceWorld, approaching the performance of state-of-the-art closed-source systems while reducing per-step latency by 60%.

Abstract

Existing memory systems for embodied agents typically inject retrieved memory as static context at episode start, a paradigm we term Ahead-of-time Monolithic Memory Injection (AMMI). However, this static design quickly becomes misaligned with the agent's evolving state and may degrade lightweight executors below the no-memory baseline. To address this, we propose MemCompiler, which reframes memory utilization as State-Conditioned Memory Compilation. A learned Memory Compiler reads a structured Brief State capturing the agent's current execution state and dynamically selects and compiles only relevant memory into executable guidance. This guidance is delivered through a text channel and a latent Soft-Mem channel that preserves perceptual information not expressible in text. Across Alf World, EmbodiedBench, and ScienceWorld, MemCompiler consistently improves over no-memory across open-source backbones (up to +129%), matches or approaches frontier closed-source systems, and reduces per-step latency by 60%, demonstrating that state-aware memory compilation improves both effectiveness and efficiency.

Problem

• embodied agents
• memory injection
• state alignment
• static context
• memory efficiency

Innovation

• State-Conditioned Memory Compilation
• MemCompiler
• Soft-Mem Channel
• Embodied Agents
• Dynamic Memory Selection