M$^2$: Dual-Memory Augmentation for Long-Horizon Web Agents via Trajectory Summarization and Insight Retrieval

TL;DR

The paper introduces M$^2$, a memory-augmented framework for web agents that enhances long-horizon task performance by compressing history and retrieving insights, reducing computational costs and improving reasoning.

Contribution

M$^2$ is a training-free, dual-memory system that combines trajectory summarization and insight retrieval to improve web agent efficiency and effectiveness in complex tasks.

Findings

Up to 19.6% success rate increase on benchmarks.

58.7% reduction in token usage for large models.

Significant accuracy improvements with lower computational costs.

Abstract

Multimodal Large Language Models (MLLMs) based agents have demonstrated remarkable potential in autonomous web navigation. However, handling long-horizon tasks remains a critical bottleneck. Prevailing strategies often rely heavily on extensive data collection and model training, yet still struggle with high computational costs and insufficient reasoning capabilities when facing complex, long-horizon scenarios. To address this, we propose M$^2$, a training-free, memory-augmented framework designed to optimize context efficiency and decision-making robustness. Our approach incorporates a dual-tier memory mechanism that synergizes Dynamic Trajectory Summarization (Internal Memory) to compress verbose interaction history into concise state updates, and Insight Retrieval Augmentation (External Memory) to guide the agent with actionable guidelines retrieved from an offline insight bank. Extensive evaluations across WebVoyager and OnlineMind2Web demonstrate that M$^2$ consistently surpasses baselines, yielding up to a 19.6% success rate increase and 58.7% token reduction for Qwen3-VL-32B, while proprietary models like Claude achieve accuracy gains up to 12.5% alongside significantly lower computational overhead.