Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

TiMem: Temporal-Hierarchical Memory Consolidation for Long-Horizon Conversational Agents

Kai Li, Xuanqing Yu, Ziyi Ni, Yi Zeng, Yao Xu, Zheqing Zhang, Xin Li, Jitao Sang, Xiaogang Duan, Xuelei Wang, Chengbao Liu, Jie Tan

TL;DR

TiMem introduces a temporal--hierarchical memory framework for long-horizon conversational agents by constructing a five-level Temporal Memory Tree (TMT) that progressively consolidates memories from raw observations to a stable persona. Consolidation is instruction-guided and model-agnostic, enabling memory abstraction across segment, session, day, week, and profile levels without fine-tuning. A complexity-aware recall pipeline—consisting of a planner, hierarchical recall, and recall gating—adapts retrieval to query difficulty, balancing precision and efficiency. Empirical results on LoCoMo and LongMemEval-S show state-of-the-art accuracy and substantial memory-usage reductions, supported by manifold analyses demonstrating improved persona separation and noise suppression. TiMem offers a practical, interpretable foundation for scalable, grounded long-horizon memory in conversational AI.

Abstract

Long-horizon conversational agents have to manage ever-growing interaction histories that quickly exceed the finite context windows of large language models (LLMs). Existing memory frameworks provide limited support for temporally structured information across hierarchical levels, often leading to fragmented memories and unstable long-horizon personalization. We present TiMem, a temporal--hierarchical memory framework that organizes conversations through a Temporal Memory Tree (TMT), enabling systematic memory consolidation from raw conversational observations to progressively abstracted persona representations. TiMem is characterized by three core properties: (1) temporal--hierarchical organization through TMT; (2) semantic-guided consolidation that enables memory integration across hierarchical levels without fine-tuning; and (3) complexity-aware memory recall that balances precision and efficiency across queries of varying complexity. Under a consistent evaluation setup, TiMem achieves state-of-the-art accuracy on both benchmarks, reaching 75.30% on LoCoMo and 76.88% on LongMemEval-S. It outperforms all evaluated baselines while reducing the recalled memory length by 52.20% on LoCoMo. Manifold analysis indicates clear persona separation on LoCoMo and reduced dispersion on LongMemEval-S. Overall, TiMem treats temporal continuity as a first-class organizing principle for long-horizon memory in conversational agents.

TiMem: Temporal-Hierarchical Memory Consolidation for Long-Horizon Conversational Agents

TL;DR

TiMem introduces a temporal--hierarchical memory framework for long-horizon conversational agents by constructing a five-level Temporal Memory Tree (TMT) that progressively consolidates memories from raw observations to a stable persona. Consolidation is instruction-guided and model-agnostic, enabling memory abstraction across segment, session, day, week, and profile levels without fine-tuning. A complexity-aware recall pipeline—consisting of a planner, hierarchical recall, and recall gating—adapts retrieval to query difficulty, balancing precision and efficiency. Empirical results on LoCoMo and LongMemEval-S show state-of-the-art accuracy and substantial memory-usage reductions, supported by manifold analyses demonstrating improved persona separation and noise suppression. TiMem offers a practical, interpretable foundation for scalable, grounded long-horizon memory in conversational AI.

Abstract

Long-horizon conversational agents have to manage ever-growing interaction histories that quickly exceed the finite context windows of large language models (LLMs). Existing memory frameworks provide limited support for temporally structured information across hierarchical levels, often leading to fragmented memories and unstable long-horizon personalization. We present TiMem, a temporal--hierarchical memory framework that organizes conversations through a Temporal Memory Tree (TMT), enabling systematic memory consolidation from raw conversational observations to progressively abstracted persona representations. TiMem is characterized by three core properties: (1) temporal--hierarchical organization through TMT; (2) semantic-guided consolidation that enables memory integration across hierarchical levels without fine-tuning; and (3) complexity-aware memory recall that balances precision and efficiency across queries of varying complexity. Under a consistent evaluation setup, TiMem achieves state-of-the-art accuracy on both benchmarks, reaching 75.30% on LoCoMo and 76.88% on LongMemEval-S. It outperforms all evaluated baselines while reducing the recalled memory length by 52.20% on LoCoMo. Manifold analysis indicates clear persona separation on LoCoMo and reduced dispersion on LongMemEval-S. Overall, TiMem treats temporal continuity as a first-class organizing principle for long-horizon memory in conversational agents.
Paper Structure (68 sections, 9 equations, 4 figures, 10 tables)

This paper contains 68 sections, 9 equations, 4 figures, 10 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Parametric Memory Approaches.
  4. External Memory Management.
  5. Methodology
  6. Temporal Memory Tree
  7. Implementation
  8. Memory Consolidation
  9. Memory Consolidator
  10. Child Memories
  11. Historical Memories
  12. Stratified Scheduling
  13. Memory Recall
  14. Recall Planner
  15. Hierarchical Recall
  16. ...and 53 more sections

Figures (4)

  • Figure 1: TiMem framework overview. The framework organizes conversational streams through a five-level TMT, consolidating memories from factual segments to persona profiles, with adaptive memory recall guided by query complexity.
  • Figure 2: TiMem architecture overview: a five-layer TMT from level 1 segments to level 5 profiles, with a consolidation pipeline processing dialog into temporal-hierarchical memories, and a recall pipeline without fine-tuning that includes a recall planner, hierarchical recall, and a recall gating module.
  • Figure 3: UMAP visualization of memory embeddings. Left: LoCoMo exhibits 10 user groups separation through hierarchical consolidation. Right: LongMemEval-S converges toward shared persona structure through noise suppression.
  • Figure 4: Case study comparing TiMem and a non-hierarchical baseline. TiMem’s hierarchical consolidation organizes timestamped evidence into coherent chains and a persona profile, whereas the baseline recalls only isolated event records.