PersonaTrail Benchmarks Personalized Web Agents with Browsing History

Seungbin Yang, Chaewoon Ki, Dohyun Lee, Jaegul Choo, ChaeHun Park· July 24, 2026 View original

Summary

Researchers introduce PersonaTrail, a new benchmark for evaluating personalized web agents that infer user context from raw browsing histories. They also propose PACMem, a framework that uses structured memory to guide agents in personalized navigation, outperforming existing baselines.

Current methods for evaluating web agents often fall short in assessing their ability to handle underspecified user instructions, particularly when agents need to infer context from past browsing data. To address this, a new benchmark called PersonaTrail has been developed. This benchmark simulates realistic browsing scenarios, allowing for the evaluation of agents' capacity to understand user preferences and recall relevant information from their history.Accompanying PersonaTrail, a novel framework named Preference-Aware Contextual Memory (PACMem) is introduced. PACMem processes raw browsing histories into two distinct memory types: factual summaries of individual sessions and distilled patterns of recurring user behavior. During operation, agents retrieve pertinent information from these structured memories to inform and personalize their navigation decisions.Experimental results indicate that PACMem significantly improves upon existing memory-based approaches across various tasks. This advancement highlights the potential for more intelligent and user-centric web agents that can adapt dynamically to individual user needs and historical interactions.

Why it matters

This research is crucial for developing more intelligent and user-friendly AI agents that can truly understand and adapt to individual user behavior and preferences on the web. Professionals can leverage these advancements to build more personalized digital experiences and automated assistants.

How to implement this in your domain

  1. 1Explore integrating PACMem-like memory structures into existing agent architectures for enhanced personalization.
  2. 2Utilize the PersonaTrail benchmark to rigorously test and validate the personalization capabilities of your own web agents.
  3. 3Analyze user browsing data to identify recurring patterns and factual information that could inform preference and factual memories.
  4. 4Develop agent strategies that dynamically retrieve and apply contextual information from structured user histories.

Who benefits

E-commerceDigital MarketingCustomer ServiceWeb DevelopmentAI Product Development

Key takeaways

  • Existing web agent benchmarks often overlook personalization based on browsing history.
  • PersonaTrail is a new benchmark for evaluating agents in personalized, open web environments.
  • PACMem is a framework that uses structured factual and preference memories from browsing history.
  • PACMem significantly improves agent performance in personalized navigation tasks.

Original post by Seungbin Yang, Chaewoon Ki, Dohyun Lee, Jaegul Choo, ChaeHun Park

"arXiv:2607.20482v1 Announce Type: new Abstract: Recent advances in large language models have enabled web agents to autonomously execute complex tasks. In practice, users frequently provide underspecified instructions, requiring agents to infer the missing context from their raw…"

View on X

Originally posted by Seungbin Yang, Chaewoon Ki, Dohyun Lee, Jaegul Choo, ChaeHun Park on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses