MemXD: Transferring Latent Behavioral Traits Across Domains for LLM Personalization

Users express preferences across many domains, and preferences stated in one context often reveal underlying behavioral traits that can inform personalization in others. Yet existing memory systems for large language model (LLM) agents tend to preserve these preferences within the domain-specific context in which they were originally expressed. Because retrieval is typically driven by semantic or keyword similarity, behavioral traits that could otherwise transfer across domains to improve personalization may not be surfaced when the preference and the query originate from different domains. We introduce MemXD, a memory system for cross-domain personalization that separates domain-agnostic behavioral traits from factual user memories. At storage time, MemXD identifies and distills transferable behaviors into latent traits, pairs each trait with question hooks encoding the decisions it could influence, and stores latent traits separately from factual memories. At retrieval time, MemXD generates factual and latent probing questions and retrieves latent traits via cosine similarity against stored hooks, bridging the semantic gap between domain-specific queries and domain-agnostic traits. We evaluate MemXD against existing memory systems on CrossMemBench and PersonaMem Task 7 across multiple LLM backbones. On CrossMemBench, MemXD achieves 60.0% cross-domain retrieval coverage and up to 35.0% accuracy, improving over the strongest evaluated baseline by 18.5 and 21 percentage points, respectively. On PersonaMem's cross-scenario generalization task, MemXD achieves up to a 7.0 percentage-point improvement over the strongest evaluated baseline. These results suggest that explicitly separating, abstracting, and retrieving behavioral traits across domains can improve cross-domain personalization in memory-augmented LLM agents. Our code is available at: https://anonymous.4open.science/r/MemXD_submission-D4C9/