MemRec: Collaborative Memory-Augmented Agentic Recommender System
MemRec enhances recommender systems with collaborative memory, achieving a 28.98% H@1 improvement on Goodreads.
Key Findings
Methodology
MemRec decouples memory management from reasoning by using a lightweight language model (LMMem) to manage a dynamic collaborative memory graph, providing distilled high-signal contexts to a downstream large language model (LLMRec) for recommendations.
Key Results
- On the Goodreads dataset, MemRec improves H@1 by 28.98% over the strongest baseline i2Agent, demonstrating superior performance on both dense and sparse data.
- On the Yelp dataset, MemRec shows excellence across various metrics, such as a 15.77% increase in H@1 and a 7.59% increase in N@5.
- By eliminating cognitive overload, MemRec achieves a 34% relative H@1 gain on the Books dataset compared to the Naive Agent.
Significance
MemRec addresses the performance bottlenecks of traditional recommender systems under data sparsity by introducing collaborative memory and asynchronous propagation, significantly enhancing recommendation accuracy and efficiency.
Technical Contribution
MemRec introduces collaborative memory graphs and asynchronous propagation mechanisms, overcoming the limitations of existing dynamic memory methods and providing new theoretical guarantees and engineering possibilities.
Novelty
This is the first to introduce collaborative memory into recommender systems, effectively addressing data sparsity and cognitive overload with asynchronous propagation.
Limitations
- In highly dynamic data environments, update frequency may impact system performance.
- Requires substantial computational resources to maintain the collaborative memory graph.
Future Work
Future work could explore applications on larger datasets, optimize the update mechanism of collaborative memory graphs, and extend to other domains.
AI Executive Summary
The evolution of recommender systems has shifted from traditional collaborative filtering to LLM-based agentic systems. However, existing systems often handle user and item memories in isolation, overlooking collaborative signals. MemRec addresses this by introducing collaborative memory, connecting isolated semantics.
The core of MemRec is architectural decoupling, using a lightweight language model (LMMem) to manage a dynamic collaborative memory graph and passing high-signal contexts to a downstream large language model (LLMRec) for recommendations. This approach not only improves recommendation accuracy but also reduces computational costs.
Experimental results show that MemRec achieves state-of-the-art performance across four benchmark datasets, with a 28.98% improvement in H@1 on the Goodreads dataset. This demonstrates MemRec's significant advantage in handling data sparsity. Future research could further optimize the update mechanism of collaborative memory graphs and explore applications in other fields.
Deep Analysis
Background
Recommender systems have evolved from traditional collaborative filtering (e.g., Sarwar et al., 2001) to dense latent embeddings in the deep learning era (e.g., Covington et al., 2016). Recently, agentic RS powered by LLMs has emerged, utilizing semantic memory for complex reasoning.
Core Problem
Existing systems often handle user and item memories in isolation, overlooking collaborative signals. This limits performance under data sparsity, making it difficult to accurately infer user needs.
Innovation
MemRec introduces collaborative memory to connect isolated semantics, addressing data sparsity. • Uses a lightweight language model (LMMem) to manage a dynamic collaborative memory graph. • Provides high-signal contexts to a downstream large language model (LLMRec) for recommendations.
Methodology
- �� Architectural decoupling: separates memory management from reasoning. • Collaborative memory retrieval: extracts efficient collaborative memory from the memory graph. • Asynchronous collaborative propagation: dynamically updates the semantic memory graph, capturing emerging trends.
Experiments
Experiments were conducted on four benchmark datasets: Amazon Books, Goodreads, MovieTV, and Yelp. Hit Rate and NDCG were used as evaluation metrics, comparing MemRec's performance with various baseline methods.
Results
MemRec improves H@1 by 28.98% on the Goodreads dataset and by 15.77% on the Yelp dataset. By eliminating cognitive overload, it achieves a 34% relative H@1 gain on the Books dataset compared to the Naive Agent.
Applications
MemRec can be used for personalized recommendations on e-commerce platforms, content recommendations on social media, and course recommendations in online education, significantly enhancing user experience.
Limitations & Outlook
In highly dynamic data environments, update frequency may impact system performance. Requires substantial computational resources to maintain the collaborative memory graph. Future research could optimize the update mechanism.
Plain Language Accessible to non-experts
Imagine a library where MemRec is like a smart librarian who remembers each reader's preferences and recommends new books based on feedback from other readers. Traditional systems recommend books based only on a reader's borrowing history, while MemRec considers other readers' opinions, providing more comprehensive recommendations.
ELI14 Explained like you're 14
Hey kiddo! Imagine you're at your school library looking for a good book. Traditional recommendation systems are like recommending books based only on what you've borrowed before. MemRec is like a super-smart librarian who not only remembers your preferences but also considers recommendations from other classmates to help you find the perfect book! Isn't that cool?
Glossary
Collaborative Memory
A memory paradigm that connects isolated semantics to share relational insights.
Used to improve recommender system performance under data sparsity.
Asynchronous Propagation
An update mechanism that allows the memory graph to evolve dynamically in the background.
Used to keep the collaborative memory graph up-to-date.
Large Language Model
A model capable of complex reasoning using natural language as a substrate.
Used for semantic memory management in recommender systems.
Cognitive Overload
Occurs when a system fails to extract key signals from a large amount of information.
In recommender systems, excessive context information may lead to reasoning difficulties.
Semantic Memory
A persistent state storing information in semantic form to help systems understand users.
Used for complex reasoning in recommender systems.
Open Questions Unanswered questions from this research
- 1 How to effectively apply MemRec to larger datasets?
- 2 How to optimize the update mechanism of collaborative memory graphs to improve performance?
Applications
Immediate Applications
E-commerce Personalization
By analyzing user and other users' purchasing behaviors, provide more accurate product recommendations.
Long-term Vision
Social Media Content Recommendation
Enhance user content discovery experience on social media platforms through collaborative memory.
Abstract
The evolution of recommender systems has shifted from traditional collaborative filtering to LLM-based agentic systems, which rely on semantic user and item memories to make predictions. However, existing agents maintain these memories in isolation. This overlooks crucial collaborative signals, such as user-item co-engagements and peer relationships across the community, which significantly limits their ability to uncover hidden preferences and accurately infer user needs, particularly for data-sparse users. To bridge this gap, we introduce collaborative memory, a paradigm that connects isolated semantics to enable the sharing of relational insights. Yet, naively utilizing collaborative memory causes severe context overload and introduces noise to downstream LLMs, alongside prohibitive computational costs. To resolve this, we propose MemRec, a framework that architecturally decouples memory management from reasoning. MemRec introduces a dedicated, lightweight language model (LM_Mem) to efficiently manage and synthesize a dynamic collaborative memory graph in the background. It provides only distilled, high-signal contexts to a downstream, heavyweight large language model (LLM_Rec) for the final recommendation. Extensive experiments on four benchmarks demonstrate that MemRec achieves state-of-the-art performance. Code: https://github.com/rutgerswiselab/memrec and Homepage: https://memrec.weixinchen.com/