Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Choosing How to Remember: Adaptive Memory Structures for LLM Agents

Mingfei Lu, Mengjia Wu, Feng Liu, Jiawei Xu, Weikai Li, Haoyang Wang, Zhengdong Hu, Ying Ding, Yizhou Sun, Jie Lu, Yi Zhang

TL;DR

A unified framework, FluxMem, is proposed that enables adaptive memory organization for LLM agents with multiple complementary memory structures, and explicitly learns to select among these structures based on interaction-level features, using offline supervision derived from downstream response quality and memory utilization.

Abstract

Memory is critical for enabling large language model (LLM) based agents to maintain coherent behavior over long-horizon interactions. However, existing agent memory systems suffer from two key gaps: they rely on a one-size-fits-all memory structure and do not model memory structure selection as a context-adaptive decision, limiting their ability to handle heterogeneous interaction patterns and resulting in suboptimal performance. We propose a unified framework, FluxMem, that enables adaptive memory organization for LLM agents. Our framework equips agents with multiple complementary memory structures. It explicitly learns to select among these structures based on interaction-level features, using offline supervision derived from downstream response quality and memory utilization. To support robust long-horizon memory evolution, we further introduce a three-level memory hierarchy and a Beta Mixture Model-based probabilistic gate for distribution-aware memory fusion, replacing brittle similarity thresholds. Experiments on two long-horizon benchmarks, PERSONAMEM and LoCoMo, demonstrate that our method achieves average improvements of 9.18% and 6.14%.

Choosing How to Remember: Adaptive Memory Structures for LLM Agents

TL;DR

A unified framework, FluxMem, is proposed that enables adaptive memory organization for LLM agents with multiple complementary memory structures, and explicitly learns to select among these structures based on interaction-level features, using offline supervision derived from downstream response quality and memory utilization.

Abstract

Memory is critical for enabling large language model (LLM) based agents to maintain coherent behavior over long-horizon interactions. However, existing agent memory systems suffer from two key gaps: they rely on a one-size-fits-all memory structure and do not model memory structure selection as a context-adaptive decision, limiting their ability to handle heterogeneous interaction patterns and resulting in suboptimal performance. We propose a unified framework, FluxMem, that enables adaptive memory organization for LLM agents. Our framework equips agents with multiple complementary memory structures. It explicitly learns to select among these structures based on interaction-level features, using offline supervision derived from downstream response quality and memory utilization. To support robust long-horizon memory evolution, we further introduce a three-level memory hierarchy and a Beta Mixture Model-based probabilistic gate for distribution-aware memory fusion, replacing brittle similarity thresholds. Experiments on two long-horizon benchmarks, PERSONAMEM and LoCoMo, demonstrate that our method achieves average improvements of 9.18% and 6.14%.
Paper Structure (58 sections, 26 equations, 11 figures, 6 tables, 3 algorithms)

This paper contains 58 sections, 26 equations, 11 figures, 6 tables, 3 algorithms.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Formalizaion
  5. Memory Granularity and Hierarchy.
  6. Memory Write and Consolidation Pipeline.
  7. Structured Episodic Memory.
  8. Query and Response Generation.
  9. Three-Layer Memory Hierarchy
  10. Short-Term Interaction Memory.
  11. Mid-Term Episodic Memory.
  12. Long-Term Semantic Memory.
  13. Multi-Structure Organization and Retrieval
  14. Linear Memory.
  15. Graph Memory.
  16. ...and 43 more sections

Figures (11)

  • Figure 1: User–agent interactions exhibit diverse structural patterns, including relational, temporal, and topical variations. Prior methods rely on a single memory structure and struggle across interaction types. Our approach dynamically selects memory structures based on the conversation, enabling robust performance.
  • Figure 2: Overall architecture of FluxMem. The framework consists of three memory layers: STIM for buffering recent context, MTEM for structure-aware episodic storage, and LTSM for long-term semantic consolidation. The central part depicts the main workflow, including memory writing, structure selection, and query-time retrieval across memory layers.
  • Figure 3: Parameter experiments results.
  • Figure 4: Time-dependent query case.
  • Figure 5: Complex-relation case.
  • ...and 6 more figures