WebCoach: Self-Evolving Web Agents with Cross-Session Memory Guidance

2026-07-25 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

WebCoach is a model-agnostic, self-evolving framework designed to enhance web browsing agents with persistent, cross-session memory, improving long-term planning and continual learning without retraining. It comprises a WebCondenser, which standardizes raw navigation logs into concise summaries; an External Memory Store (EMS), organizing complete trajectories as episodic experiences; and a Coach, which retrieves relevant experiences to inject task-specific advice. Evaluated on the WebVoyager benchmark, WebCoach consistently improved performance across three LLM backbones. For instance, a 38B model saw its task success rate increase from 47% to 61%, a 14-point gain, while maintaining or reducing average steps. Smaller base models with WebCoach achieved performance comparable to the same web agent using GPT-4o. The framework also demonstrated that self-generated experiences yield more transferable knowledge than externally seeded memories, and its benefits are strongly correlated with model scale, with significant gains on complex sites like Apple, ArXiv, and BBC News.

Key takeaway

For AI Engineers developing web agents, integrating a memory-centric framework like WebCoach can significantly boost robustness and efficiency. You should consider implementing persistent cross-session memory to enable agents to learn from past successes and failures, reducing repetitive errors and steps. This approach allows your agents to self-evolve without costly retraining, potentially achieving GPT-4o-level performance with smaller open-source models, especially for complex, multi-step web tasks.

Key insights

WebCoach augments web agents with persistent, cross-session memory for self-evolution and improved task success without retraining.

Principles

• Agents learn best from their own trajectories.
• Memory guidance is most valuable at partial competence.
• Decoupled modules enable model-agnostic improvements.

Method

WebCoach uses a WebCondenser to summarize logs, an External Memory Store for episodic experiences, and a Coach to retrieve and inject advice into the agent's prompt.

In practice

• Implement a vector database for episodic memory.
• Use a small LLM for log summarization.
• Seed memory with high-quality past episodes.

Topics

• Web Agents
• Multimodal LLMs
• Cross-Session Memory
• Continual Learning
• WebVoyager Benchmark
• Retrieval-Augmented Generation

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

Related on AIssential

• Rosetta Memory: Adaptive Memory for Cross-LLM Agents — Rosetta Memory enables cross-LLM agents to share and adapt memory effectively, shifting from LLM-centric to memory-centric design.
• rohitg00 / agentmemory — Agentmemory provides persistent, searchable memory for AI coding agents, enhancing efficiency and reducing redundant explanations.
• From Storage to Experience: A Survey on the Evolution of LLM Agent Memory Mechanisms — LLM agent memory evolves through stages of storage, reflection, and experience, driven by consistency, dynamic environments, and continual learning.
• Your AI Agent Has a Memory Problem — AI agents require robust memory systems to retain context and improve performance across sessions.
• Shared Selective Persistent Memory for Agentic LLM Systems — Selective persistent memory for agentic LLMs improves performance by retaining reusable context and discarding session-specific reasoning.
• Mem-$π$: Adaptive Memory through Learning When and What to Generate — Mem-$π$ enables LLM agents to generate context-specific guidance adaptively, outperforming static retrieval methods.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.