CrossMemBench: A Memory Benchmark for Cross-Domain Preference Transfer and Conflict Preservation

As memory-augmented LLM agents evolve toward personalized assistants, they must track user preferences that span functionally distinct domains of life, from health and finance to travel and education. A preference expressed in one domain can often inform decisions in another, yet no existing memory benchmarks offer a dedicated evaluation of whether latent preferences can transfer across domains and whether opposing preferences from different domains can be jointly surfaced to inform a single decision. We introduce CrossMemBench, a memory benchmark that targets these two capabilities through dedicated metrics, Cross-domain Memory Recall Transfer (CMRT) and Dual-Constraint Acknowledgement (DCA). We evaluate six memory baselines across five distinct life domains and 40 user profiles under increasing distractor memory pools. Through a fine-grained retrieval-versus-accuracy decomposition, we find that retrieval is the primary bottleneck across both tasks. Evaluated systems retrieve the required cross-domain memory in fewer than 45% of CMRT queries, and dual-constraint failures arise predominantly from partial retrieval, where systems surface the in-domain preference but lose the opposing cross-domain constraint. We further show that augmenting memories with domain-agnostic trait abstractions substantially raises cross-domain transfer accuracy, surpassing all evaluated memory systems, with even a weaker model using such abstractions outperforming a stronger model without them. These results suggest that effective cross-domain personalization requires both trait-level memory representations and retrieval mechanisms that prioritize decision relevance.