PersonaTrail Benchmarks Web Agents on Raw Browsing History Instead of Clean Preference Lists

arXiv 2607.20482·medium signal

Yang, Ki, Lee, Choo, and Park from KAIST target a gap in personalization benchmarks: existing suites either give agents fully explicit prompts or hand them pre-abstracted user profiles, whereas real users give underspecified instructions and the context lives in messy browsing history. PersonaTrail evaluates agents in a managed open web environment using realistic browsing trajectories as the user history, testing whether an agent can infer preferences and recall facts from past sessions. Their PACMem baseline decomposes raw history into factual memories (per-session summaries) and preference memories (recurring behavioral patterns), retrieving from both at inference time and consistently outperforming prior memory-based approaches on both task types.