Title: PersonaTree: A Structured Lifecycle Memory Framework for Person Understanding in LLM Agents

Abstract:

For long-term interactions, persistent LLM agents depend on memory systems that explicitly facilitate the development of person understanding. While current agent memory approaches prioritize the retention and retrieval of information, they often overlook the mechanisms by which accumulated interaction data is distilled into coherent person profiles. We conceptualize this distillation process as schema formation, a method wherein contextualized evidence is abstracted into reusable patterns and stable, high-level claims about individuals.

To operationalize this perspective, we present PersonaTree, a structured lifecycle memory framework organized around a three-tier persona tree. This architecture provides explicit pathways linking evidence to claims. PersonaTree manages this structure through conservative writing practices, consolidation guided by confidence levels, and query-conditioned path retrieval, ensuring that only the necessary depth of evidence is returned for each specific query.

Evaluated across six benchmarks focusing on person understanding and persistent memory, utilizing three distinct answer backbones, PersonaTree achieved the top rank in 12 out of 18 compact scores and placed in the top two in 16 different settings. Ablation studies further demonstrate that the hierarchical structure enhances abstract person understanding on the KnowMe benchmark, while support path retrieval improves alignment with RealPref under similar context constraints.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC