Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AMEM4Rec: Leveraging Cross-User Similarity for Memory Evolution in Agentic LLM Recommenders

Minh-Duc Nguyen, Hai-Dang Kieu, Dung D. Le

TL;DR

AMEM4Rec tackles the gap in LLM-based recommenders by explicitly modeling collaborative filtering signals through cross-user memory evolution. It builds a global, evolving pool of textual memories derived from user histories, links similar memories, and iteratively refines them to capture implicit CF patterns without pretrained CF models. The approach yields consistent gains over state-of-the-art baselines on four real-world datasets, especially under sparse interaction conditions, and demonstrates robustness across multiple LLM backbones. The work advances memory-augmented, agentic recommendation by enabling end-to-end collaborative pattern learning with practical considerations for prompting and retrieval.

Abstract

Agentic systems powered by Large Language Models (LLMs) have shown strong potential in recommender systems but remain hindered by several challenges. Fine-tuning LLMs is parameter-inefficient, and prompt-based agentic reasoning is limited by context length and hallucination risk. Moreover, existing agentic recommendation systems predominantly leverages semantic knowledge while neglecting the collaborative filtering (CF) signals essential for implicit preference modeling. To address these limitations, we propose AMEM4Rec, an agentic LLM-based recommender that learns collaborative signals in an end-to-end manner through cross-user memory evolution. AMEM4Rec stores abstract user behavior patterns from user histories in a global memory pool. Within this pool, memories are linked to similar existing ones and iteratively evolved to reinforce shared cross-user patterns, enabling the system to become aware of CF signals without relying on a pre-trained CF model. Extensive experiments on Amazon and MIND datasets show that AMEM4Rec consistently outperforms state-of-the-art LLM-based recommenders, demonstrating the effectiveness of evolving memory-guided collaborative filtering.

AMEM4Rec: Leveraging Cross-User Similarity for Memory Evolution in Agentic LLM Recommenders

TL;DR

AMEM4Rec tackles the gap in LLM-based recommenders by explicitly modeling collaborative filtering signals through cross-user memory evolution. It builds a global, evolving pool of textual memories derived from user histories, links similar memories, and iteratively refines them to capture implicit CF patterns without pretrained CF models. The approach yields consistent gains over state-of-the-art baselines on four real-world datasets, especially under sparse interaction conditions, and demonstrates robustness across multiple LLM backbones. The work advances memory-augmented, agentic recommendation by enabling end-to-end collaborative pattern learning with practical considerations for prompting and retrieval.

Abstract

Agentic systems powered by Large Language Models (LLMs) have shown strong potential in recommender systems but remain hindered by several challenges. Fine-tuning LLMs is parameter-inefficient, and prompt-based agentic reasoning is limited by context length and hallucination risk. Moreover, existing agentic recommendation systems predominantly leverages semantic knowledge while neglecting the collaborative filtering (CF) signals essential for implicit preference modeling. To address these limitations, we propose AMEM4Rec, an agentic LLM-based recommender that learns collaborative signals in an end-to-end manner through cross-user memory evolution. AMEM4Rec stores abstract user behavior patterns from user histories in a global memory pool. Within this pool, memories are linked to similar existing ones and iteratively evolved to reinforce shared cross-user patterns, enabling the system to become aware of CF signals without relying on a pre-trained CF model. Extensive experiments on Amazon and MIND datasets show that AMEM4Rec consistently outperforms state-of-the-art LLM-based recommenders, demonstrating the effectiveness of evolving memory-guided collaborative filtering.
Paper Structure (28 sections, 9 equations, 5 figures, 4 tables, 1 algorithm)

This paper contains 28 sections, 9 equations, 5 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. LLM-based Recommender Systems
  4. Collaborative Filtering Integration with LLMs
  5. Agentic and Memory-based Recommender Systems
  6. Methodology
  7. Preliminaries
  8. Problem Formulation
  9. Collaborative Filtering
  10. Learning Space in Agentic Recommendation
  11. Memory Definition
  12. Link Memory
  13. Memory Evolution
  14. Memory-Augmented Ranking
  15. Experiments
  16. ...and 13 more sections

Figures (5)

  • Figure 1: Overview of the AMEM4Rec framework showing three training stages: (1) Memory Creation from user interaction histories, (2) Memory Linking via dual validators (similarity + semantic), and (3) Memory Evolution through iterative updates. During inference, retrieved memories augment the LLM agent for personalized ranking.
  • Figure 2: Illustration of extracted memory components: Behavior Explanation and Pattern Description.
  • Figure 3: Prompt template for updating memory
  • Figure 4: Ablation study visualizations: (a) t-SNE embedding comparison between models with and without memory evolution mechanism, and (b) distribution comparison between full model and variant without the semantic validator.
  • Figure 5: Hyperparameter sensitivity analysis on Video Games dataset. Left: window size $w$ affects pattern granularity (optimal: $w=3$); Middle: number of retrieved memories in linking stage $|L|$ (optimal: $|L|=5$); Right: number of memories used in ranking stage $K$ (optimal: $K=5$).