Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

FadeMem: Biologically-Inspired Forgetting for Efficient Agent Memory

Lei Wei, Xu Dong, Xiao Peng, Niantao Xie, Bin Wang

TL;DR

FadeMem tackles memory scalability in autonomous LLM agents by introducing biologically-inspired forgetting through a dual-layer memory and adaptive decay driven by semantic relevance, access frequency, and temporal patterns. The approach combines LLM-guided conflict resolution and intelligent memory fusion to consolidate related memories while fading irrelevant details, yielding substantial storage reductions without sacrificing multi-hop reasoning and retrieval. Across MSC, LoCoMo, and LTI-Bench, FadeMem shows improved retention of critical facts (e.g., ~82% retention with ~55% storage), robust conflict handling, and strong cross-dataset performance, highlighting the practical value of selective forgetting. By emulating human forgetting curves and consolidation dynamics, FadeMem offers a scalable framework for long-term agent memory with dynamic capacity and retrieval accuracy.

Abstract

Large language models deployed as autonomous agents face critical memory limitations, lacking selective forgetting mechanisms that lead to either catastrophic forgetting at context boundaries or information overload within them. While human memory naturally balances retention and forgetting through adaptive decay processes, current AI systems employ binary retention strategies that preserve everything or lose it entirely. We propose FadeMem, a biologically-inspired agent memory architecture that incorporates active forgetting mechanisms mirroring human cognitive efficiency. FadeMem implements differential decay rates across a dual-layer memory hierarchy, where retention is governed by adaptive exponential decay functions modulated by semantic relevance, access frequency, and temporal patterns. Through LLM-guided conflict resolution and intelligent memory fusion, our system consolidates related information while allowing irrelevant details to fade. Experiments on Multi-Session Chat, LoCoMo, and LTI-Bench demonstrate superior multi-hop reasoning and retrieval with 45\% storage reduction, validating the effectiveness of biologically-inspired forgetting in agent memory systems.

FadeMem: Biologically-Inspired Forgetting for Efficient Agent Memory

TL;DR

FadeMem tackles memory scalability in autonomous LLM agents by introducing biologically-inspired forgetting through a dual-layer memory and adaptive decay driven by semantic relevance, access frequency, and temporal patterns. The approach combines LLM-guided conflict resolution and intelligent memory fusion to consolidate related memories while fading irrelevant details, yielding substantial storage reductions without sacrificing multi-hop reasoning and retrieval. Across MSC, LoCoMo, and LTI-Bench, FadeMem shows improved retention of critical facts (e.g., ~82% retention with ~55% storage), robust conflict handling, and strong cross-dataset performance, highlighting the practical value of selective forgetting. By emulating human forgetting curves and consolidation dynamics, FadeMem offers a scalable framework for long-term agent memory with dynamic capacity and retrieval accuracy.

Abstract

Large language models deployed as autonomous agents face critical memory limitations, lacking selective forgetting mechanisms that lead to either catastrophic forgetting at context boundaries or information overload within them. While human memory naturally balances retention and forgetting through adaptive decay processes, current AI systems employ binary retention strategies that preserve everything or lose it entirely. We propose FadeMem, a biologically-inspired agent memory architecture that incorporates active forgetting mechanisms mirroring human cognitive efficiency. FadeMem implements differential decay rates across a dual-layer memory hierarchy, where retention is governed by adaptive exponential decay functions modulated by semantic relevance, access frequency, and temporal patterns. Through LLM-guided conflict resolution and intelligent memory fusion, our system consolidates related information while allowing irrelevant details to fade. Experiments on Multi-Session Chat, LoCoMo, and LTI-Bench demonstrate superior multi-hop reasoning and retrieval with 45\% storage reduction, validating the effectiveness of biologically-inspired forgetting in agent memory systems.
Paper Structure (13 sections, 15 equations, 1 figure, 3 tables)

This paper contains 13 sections, 15 equations, 1 figure, 3 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Dual-Layer Memory Architecture with Differential Forgetting
  4. Biologically-Inspired Forgetting Curves
  5. Memory Conflict Resolution
  6. Adaptive Memory Fusion
  7. Experiments
  8. Experimental Setup
  9. Memory Retention and Forgetting Dynamics
  10. Conflict Resolution Performance
  11. Cross-Dataset Evaluation
  12. Ablation Study
  13. Conclusion

Figures (1)

  • Figure 1: Ablation study results on LoCoMo across different task types. Each bar shows F1 scores when removing specific components compared to the full model.