Source-linked AI summary
AMEM4Rec: Leveraging Cross-User Similarity for Memory Evolution in Agentic LLM Recommenders
Minh-Duc Nguyen, Hai-Dang Kieu, Dung D. Le
TL;DR
LLM recommenders face context, efficiency, and collaborative-filtering limitations, especially when implicit interaction signals are not explicitly modeled. AMEM4Rec addresses these issues by evolving cross-user behavior memories in a global pool, and experiments report consistent gains over strong LLM-based baselines. Its deployment remains bounded by prompting and embedding quality, API costs, and extremely sparse data.
Problem
LLM recommenders face context-length, computational-efficiency, and collaborative-filtering limitations when modeling implicit user–item preferences.
Method
AMEM4Rec creates abstract behavior memories, links similar entries, and iteratively evolves them to aggregate cross-user collaborative patterns without pretrained CF models.
Results
AMEM4Rec consistently outperforms strong prompt-based, generative, and hybrid LLM recommenders across Amazon and MIND datasets.
Takeaways & Limitations
The results support memory evolution as an end-to-end way to incorporate collaborative filtering signals into agentic LLM recommendation.
Takeaways & Limitations
Performance depends on prompting and embedding quality, incurs API costs, and may be limited on extremely sparse datasets.
Abstract
from arXiv · showhide
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.
1 Introduction
LLM recommenders offer semantic and generative capabilities, but prompt-based and fine-tuned approaches face context, noise, cost, and update constraints while often neglecting collaborative filtering. AMEM4Rec addresses this gap by evolving cross-user behavior memories to capture collaborative signals end to end.
- Prompt-based recommenders can suffer from long contexts and information noise, while fine-tuning is computationally expensive for large models and frequent updates.
- Existing LLM-based approaches often fail to explicitly model implicit collaborative filtering signals from user–item interactions.
- AMEM4Rec stores user behavior patterns as memories, links similar behaviors, and continually refines entries to aggregate recurring cross-user interaction patterns.
- AMEM4Rec frames semantic collaborative filtering as textual memory representations that aggregate implicit signals across users rather than relying on explicit co-occurrence matrices.
- Across four real-world datasets, AMEM4Rec consistently outperforms state-of-the-art LLM-based recommenders across metrics, especially in sparse interaction scenarios.
2 Related Work
Related work has used LLMs for semantic representations and attempted to inject collaborative knowledge, while agentic systems use profiles and memories for personalization. AMEM4Rec targets the remaining gap: end-to-end cross-user memory evolution without pretrained CF models.
- LLM-based recommenders commonly use semantic embeddings or generated side information to enrich user and item representations.
- Methods bridging CF and LLMs often transfer collaborative knowledge from pretrained CF recommenders, introducing dependence on those models.
- Agentic recommenders use multi-step reasoning, profiles, and memories to simulate user behavior and support personalized recommendations.
- Existing memory designs rarely aggregate cross-user collaborative patterns end to end without pretrained CF models.
- AMEM4Rec introduces memory evolution tailored to cross-user collaborative pattern aggregation for sparse and dynamic recommendation scenarios.
3 Methodology
AMEM4Rec trains a memory pool through memory creation, linking and validation, and memory evolution, then uses the trained memories to augment agentic ranking. This pipeline embeds collaborative filtering signals into memory-guided recommendation.
- The framework creates memories from user histories, links and validates them, evolves the memory pool, and uses the trained system for inference-time ranking.
3.1 Preliminaries
The paper formulates recommendation as memory-augmented re-ranking: a base recommender supplies candidates, and an LLM agent uses user history and evolving memory to refine their order. The agent combines a core LLM with a trainable memory component rather than fine-tuning the LLM parameters.
- The proposed framework re-ranks candidate items retrieved by a base recommender using user behavior memory.
- The recommendation objective is to refine and reorder a candidate set into a more accurate personalized ranking.
- LLMs primarily provide semantic understanding but struggle to learn implicit interaction patterns from sparse data without explicit guidance.
- AMEM4Rec learns collaborative signals through cross-user memory evolution without relying on pretrained CF models.
- The model combines a core LLM containing semantic knowledge with a memory model trained on recommendation data and updated as user behavior changes.
- Extracted memories contain behavior explanations and pattern descriptions that represent user intent and recurring interaction structures.
3.2 Memory Definition
AMEM4Rec represents user behavior as textual memories extracted from sliding windows of interaction histories. These memories combine interpretable behavior explanations with recurring cross-user patterns and are stored in a shared pool for retrieval.
- Memory construction: Each memory fragment is a user behavior pattern extracted from a sliding window over a user’s temporally ordered interaction history.The window contains item titles and categories, which an LLM converts into a structured pattern description.
- Memory representation: The LLM-generated pattern description contains a behavior explanation and a pattern description.The behavior explanation captures motivations, while the pattern representation summarizes recurring interaction structures.
- Memory representation: Behavior explanations encode interpretable user intent, while pattern representations encode abstract collaborative patterns shared across users.Together, these components represent both individual motivations and recurring behavioral regularities.
- Memory storage: The generated textual knowledge is encoded into a shared embedding space to support later memory retrieval.
- Memory storage: Memory creation iterates over all users and stores each memory as a pattern–embedding pair in the global memory pool.The memory is defined as m_k = (p_k, e_k).
3.3 Link Memory
AMEM4Rec links each new memory to similar entries in a shared pool, then uses similarity distributions and semantic validation to decide whether memories should be updated or stored. This soft, distribution-aware process supports selective cross-user memory linking while limiting redundancy and false knowledge.
- Candidate retrieval: A new memory is compared with every existing memory using cosine similarity, and the top-k most similar entries become linking candidates.The comparison uses the new memory embedding and existing memory embeddings.
- Similarity validation: The similarity validator uses multiple strategies instead of one hard threshold to control memory updates and storage more flexibly.
- Similarity validation: Two thresholds, τ_low and τ_high, divide similarity scores into three operational zones, while score proportions characterize the retrieved neighborhood.The maximum similarity and the proportions of high, medium, and low scores inform the memory-management policy.
- Memory management: When similarity is low, the new memory is stored; when highly similar memories are concentrated, existing memories are updated to avoid redundancy.Intermediate cases use distribution-aware decisions, while dissimilar memories are stored to prevent false knowledge.
- Semantic validation: A semantic validator selects memories for updating by analyzing textual pattern content rather than relying only on cosine similarity.This validator operates after similarity-based retrieval of candidate memories.
- Candidate linking: The linking process collects candidate memories in L, which represents the entries linked to the new memory.
3.4 Memory Evolution
Memory evolution uses a separate agent to select linked memories and update them with information from a new memory. The resulting replacements allow memories to incorporate historical patterns and emerging trends across users.
- Update selection: A separate agent selects which linked memory should be updated using the new memory and the candidate set.The candidates are generated during the linking step and passed to the agent for an update decision.
- Memory replacement: The updated memory m* directly replaces the selected memory m_i in the memory pool.
- Memory evolution: Figure 3 presents the prompt template used to update memory.
3.5 Memory-Augmented Ranking
During ranking, AMEM4Rec prompts the agent to reorder candidate items using the user’s history, relevant memory fragments, and the candidate list. Relevant memories are retrieved by embedding the recent history and selecting items by cosine similarity.
- Ranking: The ranking agent reorders candidate items to identify the most suitable next item.
- Ranking inputs: Ranking uses the user’s historical items, relevant memory fragments, and the candidate list as input.
- Memory retrieval: Relevant memories are retrieved by encoding the user’s recent titles and categories and selecting memories with the highest cosine similarity.
4 Experiments
AMEM4Rec is evaluated across multiple datasets, baselines, ablations, hyperparameter settings, cold-start users, and LLM backbones. Results consistently support cross-user memory evolution, dual validation, and moderate memory settings for ranking recommendation.
- Experimental Setup: Experiments use four datasets, sampled users with more than 10 interactions, and leave-one-out evaluation, with NDCG@K at K=1,5,10 as the ranking metric.The study uses Fashion, Video Games, CDs and Vinyl, and MIND; 300 users are sampled for training and evaluation because of API cost.
- Overall Performance: AMEM4Rec consistently outperforms all baselines across datasets and NDCG@K metrics.The comparison covers ranking recommenders including LLMRank, AgentCF, and iAgent.
- Ablation Study: The complete model outperforms variants without memory evolution, similarity validation, or semantic validation.The ablation results attribute the strongest performance to combining both validators with memory evolution.
- Ablation Study: Memory evolution produces clearer embedding clusters, while removing semantic validation increases evolution counts but leaves performance inferior.The reported analysis links semantic filtering to excluding noisy or irrelevant memories from evolution.
- Hyperparameter Analysis: The best hyperparameter settings are w=3, |L|=5, and K=5; larger settings tend to reduce or destabilize performance.The window size controls pattern granularity, while |L| and K control memories used during linking and ranking.
- Further Analyses: AMEM4Rec outperforms baselines for cold-start users with 2–3 interactions and remains strong across closed-source and open-source LLM backbones.The paper attributes cold-start performance to patterns aggregated from cross-user behaviors.
5 Conclusion and Future Work
AMEM4Rec incorporates collaborative filtering through cross-user memory evolution and outperforms strong recommender baselines across Amazon and MIND datasets. The authors identify prompting, embedding quality, API cost, and extreme sparsity as remaining constraints, while proposing reinforcement-learning-based memory control for future work.
- Conclusion: AMEM4Rec creates abstract behavior memories, links similar cross-user patterns, and iteratively evolves them for end-to-end collaborative learning without pretrained CF models.Ablations identify memory evolution and link generation as important for aggregating collaborative signals and improving ranking quality, particularly under sparse interactions.
- Conclusion: AMEM4Rec consistently outperforms prompt-based, generative, and hybrid CF-knowledge baselines on Amazon Fashion, Video Games, CDs and Vinyl, and MIND.
- Limitations: Performance depends heavily on LLM prompting and embedding quality, while API costs remain a practical concern for large-scale deployment.
- Limitations: Extremely sparse datasets may limit propagation of collaborative signals through memory evolution.
- Future Work: Future work will combine textual memories with learnable parameters optimized through reinforcement learning to control retrieval and evolution more adaptively.
A Prompt Templates
The prompt templates operationalize behavior extraction, cross-user pattern linking, memory evolution, and category-level item ranking. They require concise structured JSON outputs while constraining recommendations to general behavioral and category trends rather than specific item identities.
- Behavior Pattern Extraction: Pattern descriptions may mention specific items only when they clearly represent a general sequence or co-occurrence, such as a console followed by an accessory.
- Behavior Pattern Extraction: Behavior extraction produces concise explanations of stable high-level user tendencies and concrete interaction structures such as sequences, repetition, or co-purchase.
- Collaborative Pattern Linking: Pattern-linking decisions determine whether similar behaviors should be linked, updated, or stored separately based on shared categories, preferences, trends, or distinct user segments.
- Collaborative Pattern Linking: The linking prompt returns a Boolean link decision, linked thought identifiers, and concise reasoning for the decision.
- Collaborative Memory Evolution: Memory evolution updates a candidate only when the new pattern meaningfully strengthens or refines a shared category, preference, or sequence.Otherwise, the original text is retained, and updates include revised explanations, descriptions, and brief reasoning.
- Candidate Ranking: Ranking uses recent user history and cross-user memory insights while focusing only on category-level trends and general behavioral tendencies.
- Candidate Ranking: The ranking template prioritizes consistency with history and memory, favors broader categories when evidence is limited, and returns every candidate in valid JSON.For similar items, it prefers candidates aligned with cross-user trends while requiring concise reasoning.