[arXiv]score: 0.15
PersonaTree Builds Hierarchical Long-Term User Models for Persistent Agents
June 4, 2026
PersonaTree structures agent memory as a three-level persona tree that abstracts interaction evidence into stable person-level claims via confidence-guided consolidation and query-conditioned path retrieval. It ranks first in 12 of 18 conditions across six person-understanding and persistent memory benchmarks with three answer backbones.
cs.CL
HOW THIS AFFECTS YOU
●
builderYou can use this framework to give persistent agents structured, updatable user models that go beyond flat memory retrieval for personalization tasks.
●
researcherThe schema formation framing and evidence-to-claim support paths provide a principled architecture for studying how agents should abstract long-term user understanding.