Agentic Recommender System with Hierarchical Belief-State Memory
MARS employs a three-tier memory architecture with adaptive lifecycle management, achieving 26.4% HR@1 improvement in recommendation tasks.
Key Findings
Methodology
This paper introduces MARS, modeling recommendation as a POMDP with a hierarchical belief state comprising event, preference, and profile memories. Event memory stores raw behavioral signals, preference memory maintains fine-grained, mutable preference chunks with explicit strength and evidence, and profile memory synthesizes these into a coherent natural language narrative. Six operations—extraction, reinforcement, weakening, consolidation, forgetting, and resynthesis—are governed by an LLM-based planner that adaptively schedules updates based on accumulated signals, balancing freshness and computational cost. Experiments across four InstructRec domains show that MARS outperforms baselines by 26.4% in HR@1 and 10.3% in NDCG@10, demonstrating the effectiveness of hierarchical memory and adaptive scheduling.
Key Results
- Across four datasets, MARS achieves state-of-the-art results, with HR@1 improvements ranging from 8.2% to 46.0% over the best baselines. The adaptive scheduler enhances performance in evolving scenarios, especially when user preferences shift rapidly. Ablation studies confirm that the hierarchical structure and dynamic lifecycle operations are crucial for these gains.
- Compared to models like MemRec and i2Agent, MARS consistently delivers higher recommendation accuracy and more efficient memory management, especially in sparse or dynamic environments. The results validate the hypothesis that structured, evolving memory improves personalization.
- The ablation analysis indicates that each component—hierarchical memory, lifecycle operations, and adaptive scheduling—contributes significantly to overall performance, with the combined approach yielding the best results.
Significance
This work addresses fundamental limitations in recommendation systems by integrating hierarchical, evolving memory with natural language understanding. It enables systems to better capture short-term signals and long-term preferences, providing more accurate, interpretable, and adaptable recommendations. The proposed lifecycle management framework offers a new paradigm for persistent, context-aware user modeling, with broad implications for personalized content delivery, e-commerce, and social platforms. It bridges the gap between static memory models and dynamic user behavior, paving the way for more intelligent, human-like recommendation agents.
Technical Contribution
The core technical innovation lies in the hierarchical belief state architecture combined with a full memory lifecycle governed by an LLM-based scheduler. This design allows precise control over memory evolution, enabling fine-grained credit assignment, selective forgetting, and coherent profile synthesis. Unlike prior flat or static models, MARS offers a principled approach to managing noisy behavioral signals over time, with formalized operations and adaptive decision-making, opening new avenues for scalable, interpretable recommendation systems.
Novelty
This is the first work to formalize a hierarchical belief-state model with a complete lifecycle for recommendation memory, integrating it with LLM-driven adaptive scheduling. It systematically addresses the limitations of flat memory representations and fixed update intervals, introducing a novel combination of state abstraction, lifecycle operations, and dynamic planning, setting a new standard in memory-augmented recommendation research.
Limitations
- The reliance on large pre-trained LLMs incurs high computational costs, limiting real-time deployment in resource-constrained environments.
- Preference chunk generation and merging strategies need further refinement to handle highly sparse or rapidly changing preferences accurately.
- The approach may face challenges in extremely sparse datasets or in scenarios with abrupt preference shifts, requiring more robust adaptation mechanisms.
Future Work
Future research will explore multi-modal data integration, reducing computational overhead, and extending the framework to multi-task and cross-domain recommendations. Additionally, improving preference inference accuracy and scalability in low-resource settings remains a key direction.
AI Executive Summary
In the rapidly expanding digital landscape, personalized recommendation systems are essential for connecting users with relevant content. Traditional models, including collaborative filtering and deep neural networks, have achieved significant offline success but struggle with understanding evolving user preferences and providing interpretability. The advent of large language models (LLMs) has introduced new possibilities, enabling systems to reason about user intent and item semantics through natural language. However, existing approaches often treat memory as flat and static, limiting their ability to adapt over time.
This paper introduces MARS, a novel recommendation framework that models user preferences as a hierarchical belief state, maintained through a structured memory system comprising event, preference, and profile layers. The key innovation is a full lifecycle of six operations—extraction, reinforcement, weakening, consolidation, forgetting, and resynthesis—governed by an LLM-based scheduler that adaptively manages memory updates. This design allows the system to balance memory freshness with computational efficiency, dynamically adjusting to user behavior shifts.
Extensive experiments across four diverse domains demonstrate that MARS outperforms state-of-the-art baselines, achieving an average HR@1 improvement of 26.4% and NDCG@10 increase of 10.3%. The results highlight the importance of hierarchical memory and adaptive lifecycle management in capturing complex user preferences and enhancing recommendation accuracy. The framework’s interpretability, enabled by natural language profiles, further supports explainability and user trust.
Despite these advances, challenges remain, including high computational costs and the need for more robust handling of sparse or rapidly changing preferences. Future work will focus on multi-modal data integration, reducing resource requirements, and extending the approach to broader recommendation scenarios. Overall, MARS marks a significant step toward more intelligent, adaptable, and human-like recommender systems, with promising applications in e-commerce, content platforms, and personalized services.
Deep Dive
Abstract
Memory-augmented LLM agents have advanced personalized recommendation, yet existing approaches universally adopt flat memory representations that conflate ephemeral signals with stable preferences, and none provides a complete lifecycle governing how memory should evolve. We propose MARS (Memory-Augmented Agentic Recommender System), a framework that treats recommendation as a partially observable problem and maintains a structured belief state that progressively abstracts noisy behavioral observations into a compact estimate of user preferences. MARS organizes this belief state into three tiers: event memory buffers raw signals, preference memory maintains fine-grained mutable chunks with explicit strength and evidence tracking, and profile memory distills all preferences into a coherent natural language narrative. A complete lifecycle of six operations -- extraction, reinforcement, weakening, consolidation, forgetting, and resynthesis -- is adaptively scheduled by an LLM-based planner rather than fixed-interval heuristics. Experiments on four InstructRec benchmark domains show that MARS achieves state-of-the-art performance with average improvements of 26.4% in HR@1 and 10.3% in NDCG@10 over the strongest baselines with further gains from agentic scheduling in evolving settings.