MemAgent: A cache inspired framework for augmenting conversational Web Agents with task-specific information
Abstract: Large Language Models (LLMs) have shown promise as web agents, but their current limitations hinder their widespread adoption for general users. A critical issue behind this is the misalignment between user expectations and the agent's actions due to ineffective communication leading to a lack of crucial context required for successful task completion. To address this gap, we propose MemAgent, a novel pipeline for LLM-based web agents. Inspired by caching mechanisms, MemAgent incorporates a memory component to store task-specific information. This memory bank enables the LLM agent to proactively query for supplementary context relevant to the current task, thereby reducing user interaction overhead. Our evaluations demonstrate that MemAgent significantly enhances the agent's performance and usabilities, bringing us a step closer to seamless LLM integration in web agent technologies.
Paper Type: Short
Research Area: Dialogue and Interactive Systems
Research Area Keywords: task-oriented; human-in-the-loop; knowledge augmented; embodied agents; applications; conversational modeling;
Contribution Types: Model analysis & interpretability, NLP engineering experiment
Languages Studied: English
Submission Number: 2767
Loading