PersonaTrail benchmarks web agents on raw browsing history instead of clean preference lists.

KAIST's Yang, Ki, Lee, Choo, and Park target a gap where existing personalization suites either give fully explicit prompts or hand agents pre-abstracted profiles, while real users give underspecified instructions and the context lives in messy history (arXiv 2607.20482). Their PACMem baseline decomposes raw history into factual memories (per-session summaries) and preference memories (recurring behavioral patterns), retrieving from both at inference and consistently outperforming prior memory-based approaches on both task types.