Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Inside Out: Evolving User-Centric Core Memory Trees for Long-Term Personalized Dialogue Systems

Jihao Zhao, Ding Chen, Zhaoxin Fan, Kerun Xu, Mengting Hu, Bo Tang, Feiyu Xiong, Zhiyu li

TL;DR

The paper addresses the challenge of maintaining coherent long-term personalization under unbounded interaction streams and finite context by introducing PersonaTree, a structured memory carrier built on a biopsychosocial schema. A lightweight MemListener learns to convert unstructured dialogue into executable memory operations (ADD, UPDATE, DELETE, NO_OP) to evolve the PersonaTree via process-based reinforcement learning, enabling memory compression and interpretability. At inference, PersonaTree guides fast generation, with an agentic recall mode for on-demand, retrieval-augmented detail when needed. Experiments on PersonaMem show that PersonaTree-based methods outperform full-context and other memory systems across multiple response models, and a small MemListener achieves competitive memory-operation performance against larger reasoning models. The work advances scalable, interpretable long-term personalization for dialogue systems by combining structured memory, learnable editing, and adaptive generation strategies.

Abstract

Existing long-term personalized dialogue systems struggle to reconcile unbounded interaction streams with finite context constraints, often succumbing to memory noise accumulation, reasoning degradation, and persona inconsistency. To address these challenges, this paper proposes Inside Out, a framework that utilizes a globally maintained PersonaTree as the carrier of long-term user profiling. By constraining the trunk with an initial schema and updating the branches and leaves, PersonaTree enables controllable growth, achieving memory compression while preserving consistency. Moreover, we train a lightweight MemListener via reinforcement learning with process-based rewards to produce structured, executable, and interpretable {ADD, UPDATE, DELETE, NO_OP} operations, thereby supporting the dynamic evolution of the personalized tree. During response generation, PersonaTree is directly leveraged to enhance outputs in latency-sensitive scenarios; when users require more details, the agentic mode is triggered to introduce details on-demand under the constraints of the PersonaTree. Experiments show that PersonaTree outperforms full-text concatenation and various personalized memory systems in suppressing contextual noise and maintaining persona consistency. Notably, the small MemListener model achieves memory-operation decision performance comparable to, or even surpassing, powerful reasoning models such as DeepSeek-R1-0528 and Gemini-3-Pro.

Inside Out: Evolving User-Centric Core Memory Trees for Long-Term Personalized Dialogue Systems

TL;DR

The paper addresses the challenge of maintaining coherent long-term personalization under unbounded interaction streams and finite context by introducing PersonaTree, a structured memory carrier built on a biopsychosocial schema. A lightweight MemListener learns to convert unstructured dialogue into executable memory operations (ADD, UPDATE, DELETE, NO_OP) to evolve the PersonaTree via process-based reinforcement learning, enabling memory compression and interpretability. At inference, PersonaTree guides fast generation, with an agentic recall mode for on-demand, retrieval-augmented detail when needed. Experiments on PersonaMem show that PersonaTree-based methods outperform full-context and other memory systems across multiple response models, and a small MemListener achieves competitive memory-operation performance against larger reasoning models. The work advances scalable, interpretable long-term personalization for dialogue systems by combining structured memory, learnable editing, and adaptive generation strategies.

Abstract

Existing long-term personalized dialogue systems struggle to reconcile unbounded interaction streams with finite context constraints, often succumbing to memory noise accumulation, reasoning degradation, and persona inconsistency. To address these challenges, this paper proposes Inside Out, a framework that utilizes a globally maintained PersonaTree as the carrier of long-term user profiling. By constraining the trunk with an initial schema and updating the branches and leaves, PersonaTree enables controllable growth, achieving memory compression while preserving consistency. Moreover, we train a lightweight MemListener via reinforcement learning with process-based rewards to produce structured, executable, and interpretable {ADD, UPDATE, DELETE, NO_OP} operations, thereby supporting the dynamic evolution of the personalized tree. During response generation, PersonaTree is directly leveraged to enhance outputs in latency-sensitive scenarios; when users require more details, the agentic mode is triggered to introduce details on-demand under the constraints of the PersonaTree. Experiments show that PersonaTree outperforms full-text concatenation and various personalized memory systems in suppressing contextual noise and maintaining persona consistency. Notably, the small MemListener model achieves memory-operation decision performance comparable to, or even surpassing, powerful reasoning models such as DeepSeek-R1-0528 and Gemini-3-Pro.
Paper Structure (44 sections, 6 equations, 4 figures, 10 tables)

This paper contains 44 sections, 6 equations, 4 figures, 10 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Personalization and Memory
  4. LLM Agents with External Memory
  5. The Inside Out Framework
  6. Overview Architecture
  7. Framework Pipeline.
  8. Problem Formulation.
  9. Dynamic PersonaTree Evolution
  10. PersonaTree Initialization.
  11. Iterative PersonaTree Updating.
  12. MemListener Training
  13. Training Data Synthesis.
  14. Warm-up via SFT.
  15. Alignment via Process-Reward RL.
  16. ...and 29 more sections

Figures (4)

  • Figure 1: Overview of the entire process of our Inside Out framework.
  • Figure 2: Radar-chart comparison of PersonaTree and baselines across multi-dimensional capability metrics under two memory-extraction settings (DeepSeek-R1-0528, DeepSeek-V3.1) and three response models (DeepSeek-V3.1, Longcat-Flash-Chat, DeepSeek-R1-0528)
  • Figure 3: Overall performance of memory-operation models across three response models.
  • Figure 4: Effect of dialogue chunking window size on performance across three response models.