PaperlayerView exact revision on arXiv

Source-linked AI summary

MemRec: Collaborative Memory-Augmented Agentic Recommender System

Weixin Chen, Yuhan Zhao, Jingyuan Huang, Zihe Ye, Clark Mingxuan Ju, Tong Zhao, Neil Shah, Li Chen, Yongfeng Zhang

arXiv:2601.08816v3cs.IRcs.AI

TL;DR

Existing agentic recommender systems isolate semantic memories and therefore miss collaborative relationships needed to infer preferences, particularly for data-sparse users. MemRec decouples memory management from reasoning to curate collaborative context and propagate graph updates efficiently. Across four benchmarks, it achieves state-of-the-art performance while balancing reasoning quality, computational cost, and deployment constraints.

  • Problem

    Existing agent-based recommender systems rely on isolated user and item memories, limiting their use of community signals for preference inference, particularly for data-sparse users.

  • Method

    MemRec decouples high-level reasoning from memory management, using collaborative retrieval and asynchronous propagation to curate compact context and update a dynamic graph.

  • Results

    MemRec achieves state-of-the-art performance across four benchmarks and establishes a Pareto frontier balancing reasoning quality, computational cost, and deployment constraints.

  • Takeaways & Limitations

    Collaborative memory provides a framework for sharing relational signals while controlling cognitive overload and update bottlenecks in agentic recommendation.

  • Takeaways & Limitations

    Propagation is restricted to immediate neighbors, curation rules rely on static offline domain statistics, and ceiling reasoning performance still depends on proprietary LLMs.

Abstract

from arXiv · show

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/

1 Introduction

Agentic recommender systems use semantic memories but typically isolate user and item information, missing collaborative signals that help infer preferences, especially for data-sparse users. MemRec addresses the resulting overload and update-cost challenges by decoupling memory management from reasoning and synthesizing compact collaborative context.

  • Background: Agentic recommender systems evolved from sparse ratings and dense embeddings toward semantic memories processed by LLMs.These memories retain and evolve user understanding through ongoing interactions.
  • Problem: Existing agents isolate user and item memories, preventing interaction signals and peer preferences from propagating across the broader user-item graph.This limits transfer from warm to cold items and discovery of hidden interests.
  • Problem: Naively injecting raw collaborative neighborhoods creates cognitive overload because abundant textual and structural signals obscure salient knowledge.The paper notes that this issue is validated in §3.3.
  • Problem: Naive collaborative memory also makes continuous graph evolution computationally prohibitive by requiring redundant independent LLM calls for related users and items.Each new interaction requires neighborhood contexts to remain synchronized.
  • MemRec: MemRec decouples reasoning from memory management, using a separate Memory Manager to maintain a dynamic graph and synthesize compact grounding.This architectural shift changes isolated memory into collaborative memory.
  • MemRec: Collaborative Memory Retrieval curates neighbor signals, while Asynchronous Collaborative Propagation batches self-reflection and neighbor updates with O(1) interaction complexity.Together, these mechanisms target retrieval overload and update bottlenecks without redundant independent updates.
  • Results: MemRec achieves state-of-the-art performance across four benchmarks and establishes a Pareto frontier balancing reasoning quality, computational cost, and deployment constraints.The framework supports setups ranging from cloud-native APIs to on-premise local models.

2 Methodology

MemRec builds a collaborative memory graph and separates memory management from recommendation reasoning. Its pipeline retrieves and synthesizes compact collaborative context, uses that context for grounded ranking, and asynchronously propagates updated semantic memories.

  • Motivation: MemRec addresses isolated agent memories that miss co-engagement and peer-preference signals needed to transfer interaction information and uncover hidden interests.Existing agents commonly represent user and item memories as separate semantic narratives, restricting access to broader user-item relationships.
  • Architecture: MemRec represents users and items as nodes in a unified memory graph whose edges encode interactions and derived relations.The graph enables high-order connectivity to synthesize and propagate collaborative signals beyond isolated node memories.
  • Architecture: Collaborative Memory Retrieval extracts a concise Mcollab from the graph, Grounded Reasoning ranks candidates with that context, and Asynchronous Collaborative Propagation updates semantic memories in the background.The three stages decouple reasoning from memory management while preserving evolving collaborative information.
  • Collaborative Memory Retrieval: Raw neighborhood injection can exceed context limits and bombard the reasoning model with noise, increasing hallucinations and reducing instruction adherence.The method therefore seeks a collaborative memory that preserves relevant graph information while discarding redundant or irrelevant signals.
  • Collaborative Memory Retrieval: MemRec uses a Curate-then-Synthesize strategy that prunes redundant neighbors before compressing their information into structured collaborative memory facets.The approach is conceptually guided by Information Bottleneck theory and uses domain-adaptive rules generated by LMMem for the first compression pass.
  • Grounded Reasoning: LMMem supplies the target user’s full semantic memory alongside compact neighbor representations, then synthesizes high-level facets grounded in neighbor evidence for LLMRec.Grounded reasoning combines Mcollab with the user instruction and candidate item memories to produce relevance scores and natural-language rationales.

3 Empirical Evaluation

MemRec is evaluated across four benchmark datasets and multiple research questions covering performance, architectural decoupling, deployment flexibility, ablations, robustness, and rationale quality. It achieves state-of-the-art ranking performance while improving efficiency, robustness, and rationale quality through collaborative memory and decoupled memory management.

  • Experimental Setup: Experiments use four benchmark datasets spanning diverse domains and interaction densities, with instruction-following splits from InstructRec.The datasets are Amazon Books, Amazon Goodreads, MovieTV, and Yelp.
  • Main Results: MemRec achieves state-of-the-art performance across all reported ranking metrics on four benchmarks, with all improvements over the best baseline statistically significant at p < 0.05.On Goodreads, H@1 improves by +28.98% relative to i2Agent; on Yelp, H@1 and N@5 improve by +15.77% and +7.59%.
  • Main Results: Collaborative memory methods outperform dynamic, static, and no-memory paradigms, while isolated-memory dynamic agents remain below MemRec.The reported hierarchy is Collaborative > Dynamic > Static > No Memory.
  • Architectural Impact: Architectural decoupling breaks the Naive Agent’s information bottleneck by separating memory management from reasoning and supplying the ranker with curated context.MemRec achieves a +34% relative H@1 gain on Books over the monolithic approach.
  • Flexibility and Cost: MemRec establishes a broader efficiency-performance Pareto frontier by offloading dense graph processing to asynchronous offline batches and querying distilled memory online.The analysis considers reasoning performance, sequential latency, computational cost, and deployment configurations.
  • Ablation Studies: Removing collaborative retrieval, LLM curation, or asynchronous propagation reduces H@1 by 9.9%, 5.5%, or 4.2%, respectively.The ablation results support contributions from global graph retrieval, noise-filtering curation, and dynamic collaborative updates.
  • Robustness and Quality: MemRec maintains Hit@1 of 0.491 under 30% noise injection and yields a +91.4% gain for data-sparse niche users over Vanilla.The reported analyses also show improved rationale specificity and relevance, with a hyperparameter sweet spot near k ∈{16, 32} and Nf = 7.

4 Related Works

Related work progresses from retrieval and general-purpose memory architectures to agentic recommender systems with semantic memory. Existing agentic approaches still update user or item memories in isolation rather than exploiting collaborative connectivity.

  • Memory Architectures: General-purpose memory systems support long-horizon coherence through retrieval, memory management, and reflective synthesis, but target factual or conversational domains.They do not address the specialized connectivity required for collaborative recommendation.
  • Agentic Recommenders: Agentic recommender systems have moved from stateless prompting toward explicit dynamic memory, yet updates remain confined to individual user or item silos.This leaves broader collaborative signals underused.

5 Conclusion

The paper identifies collaborative memory as a design paradigm for overcoming isolated agent memories and instantiates it through MemRec. MemRec combines decoupled reasoning and memory management to improve ranking, efficiency, and deployment flexibility, while leaving larger-scale and privacy-preserving extensions for future work.

  • Contribution: MemRec addresses isolated agent memories by connecting collaborative signals through a decoupled LLMRec reasoning and LMMem memory-management architecture.Zero-shot curation mitigates cognitive overload, while asynchronous graph propagation addresses update bottlenecks.
  • Conclusion: Across four benchmarks, MemRec improves ranking performance and establishes a Pareto frontier balancing reasoning quality, online inference cost, and deployment flexibility.The conclusion identifies web-scale graphs and privacy-preserving federated memory updates as future directions.

6 Limitations

The paper’s limitations concern the scope and adaptability of collaborative propagation, curation, and reasoning infrastructure. Its related baselines and datasets span traditional latent methods, semantic-memory agents, and four recommendation domains.

  • Limitations: Collaborative propagation is currently restricted to immediate neighbors, leaving efficient multi-hop community updates as an open problem.The limitation is motivated by computational overhead and the need to avoid introducing noise.
  • Limitations: Context-curation rules rely on static domain statistics generated offline and may require online adaptation in highly dynamic environments such as news.This bounds the current curation design’s adaptability.
  • Limitations: Ceiling reasoning performance still depends on powerful proprietary LLMs even though memory operations can be offloaded to local models.The paper identifies fully open-source stacks as future work.
  • Datasets: The evaluation uses four datasets covering e-commerce, social reading, entertainment, and local services, with domain-specific sparsity, density, recency, and contextual constraints.The datasets are Books, Goodreads, MovieTV, and Yelp.
  • Baselines: Baselines include traditional latent-embedding models and post-AgentRS approaches organized by no explicit memory, static memory, and dynamic isolated updates.Examples include LightGCN, SASRec, P5, Vanilla LLM, iAgent, i2Agent, AgentCF, and RecBot.

A.3 Implementation Details

MemRec uses diverse cloud and local model configurations, fixed reproducible hyperparameters, and a tiered neighbor representation to fit collaborative signals within a strict context budget. Domain-adaptive heuristic rules further filter collaborative subgraphs before synthesis.

  • Model Deployment: MemRec evaluates proprietary, open-weight cloud, and local open-source models across Standard, Ceiling, Cloud-OSS, Local-Qwen, and Local-Llama configurations.The local configurations use Qwen-2.5-7B-Instruct and Meta-Llama-3-8B-Instruct; cloud configurations use gpt-4o-mini, gpt-4o, and gpt-oss-120b.
  • Hardware Environment: Local-Qwen and Local-Llama experiments run on a single NVIDIA RTX A5000 GPU with 24GB VRAM.The local setup has higher latency than optimized cloud APIs, while enterprise-grade inference hardware could provide competitive speeds for 7B models.
  • Hyperparameters: The implementation fixes k = 16 neighbors, Nf = 7 synthesis facets, τ = 1800 retrieval tokens, and temperature 0.0 for reproducibility.These settings control neighbor coverage, synthesis breadth, context size, and deterministic LLM generation.
  • Neighbor Representation Strategy: Item neighbors use truncated semantic memories, whereas user neighbors use titles from their last three acted items as token-efficient proxies for immediate interests.This tiered representation preserves broad coverage within τ = 1800 without incurring prohibitive latency.
  • Collaborative Subgraph Curation: LMMem generates zero-shot, domain-specific heuristic rules from domain statistics to filter collaborative subgraphs before memory synthesis.The rules emphasize genre/theme and social signals in Books and Goodreads, while recency and categorical constraints dominate MovieTV and Yelp.
  • Cost Assumptions: The reported cloud cost estimates use Azure OpenAI Standard-tier pricing from December 2025, with local deployment treated as having negligible marginal cost.The listed rates are $2.50/$10.00 per 1M input/output tokens for gpt-4o and $0.15/$0.60 for gpt-4o-mini and gpt-oss-120b.

B.1 Memory Architectures for LLM Agents

LLM-agent memory systems evolved from external retrieval and structured memory toward explicit, dynamic semantic profiles. Yet recommendation agents largely retain isolated user or item memories rather than managing collaborative graph signals.

  • Memory Architectures: Long-horizon agents use external vector databases, Graph RAG, virtual context management, and temporal knowledge graphs to retain and structure information.Examples include RAG, Graph RAG, MemGPT, and Zep.
  • Memory Management: Dedicated memory managers and modular agent frameworks increasingly optimize storage, retrieval, tool use, and multi-dimensional memory for complex workflows.This line includes learning-based memory policies, LangChain, AutoGPT, and related modular components.
  • Agentic Recommendation Memory: Recommendation memory progressed from latent sequential states and stateless LLM prompting to explicit semantic profiles for users and items.These profiles preserve evolving user preferences or item characteristics across historical contexts.
  • Dynamic Memory: Recent recommendation agents add dynamic memory through planning, tools, or self-reflection, but updates generally remain confined to the interacting user or item.Systems such as i2Agent, RecBot, AgentCF, Agent4Rec, and RecAgent exemplify this direction.
  • Research Gap: Existing graph-based recommendation approaches typically use LLMs for feature enhancement, structure refinement, vocabulary learning, or retrieval augmentation rather than collaborative-memory management.This leaves collaborative relationships insufficiently integrated into agent memory.
  • Niche Users: MemRec’s robustness analysis connects niche-user performance with user-side fairness and long-tail recommendation challenges.The cited prior work addresses data sparsity through cross-domain transfer, fairness-aware representation learning, or neutral and uncertainty signals.

C Extended Efficiency and Modularity Analysis

The efficiency analysis reports architectural metrics across MemRec configurations and emphasizes that measured latency depends strongly on the execution environment. Deployment optimizations and larger candidate-set experiments provide additional context for interpreting efficiency and robustness.

  • Architectural Comparison: Table 5 compares configurations using H@1, N@5, sequential online latency, total token consumption, and qualitative serving cost.It distinguishes online tokens from total session tokens, which include asynchronous offline memory maintenance.
  • Deployment Optimizations: Caching synthesized collaborative contexts and streaming final outputs can reduce perceived latency in real-world deployments.Caching avoids repeated Stage-R computation for popular items, while streaming reduces time-to-first-byte.
  • Latency Interpretation: Cloud API latency comparisons require caution because the experiments used non-real-time endpoints and opaque provider factors can reverse expected model-speed rankings.The Ceiling gpt-4o configuration was measured faster than Standard gpt-4o-mini, which the authors attribute to load balancing, resource allocation, or transient network conditions.

D.2 Rationale Quality Analysis

The rationale-quality analysis evaluates how user history and collaborative memory affect specificity, relevance, and factuality. Collaborative memory improves the reported rationale dimensions while MemRec’s architecture also emphasizes token-efficient, modular operation across configurations.

  • Evaluation Protocol: GPT-4o judges Vanilla, MemRec without collaboration, and full MemRec on specificity, relevance, and factuality using 1–5 Likert ratings.The evaluation compares no memory, static user history, and collaborative memory conditions.
  • Modularity and Efficiency: MemRec’s architectural analysis separates online reasoning from offline memory maintenance and reports configurations spanning cloud and local deployment constraints.Table 5 tracks online versus total tokens, while offline maintenance incurs zero API cost when locally hosted.
  • Specificity: Collaborative memory produces a further significant specificity boost beyond user history alone (p < 0.001).The analysis attributes the gain to neighbor signals supplying specific item details for recommendations.
  • Relevance: Full collaborative context significantly increases perceived relevance (p < 0.001), whereas user history alone does not significantly improve relevance over the Base LLM (p > 0.05).The reported interpretation is that peer experiences make recommendations feel more relevant and convincing to the judge.
  • Factuality: MemRec maintains factuality above 4.0 across models and improves it slightly but significantly over the no-collaboration condition (p < 0.001).The analysis links collaborative grounding with fewer hallucinations than ungrounded generation.
  • Token Economics: Input tokens comprise nearly 80% of Standard-configuration usage, making effective cost lower than estimates based only on total token counts.Stage-R synthesis and Stage-W propagation process large input contexts into condensed outputs under asymmetric pricing.
  • Sensitivity Analysis: Hyperparameter sensitivity is consistent across H@1, H@3, H@5, NDCG@3, and NDCG@5, supporting a robust optimal region on the books-1k subset.The analysis varies neighbor count k and synthesis facets Nf.
  • Ranking Performance: Across three datasets and ranking cutoffs K = {1, 3, 5}, MemRec consistently outperforms Vanilla and Naive Agent baselines.Table 9 reports Hit Rate and NDCG metrics at the listed cutoffs.

F Prompt Templates and Contexts

The appendix documents MemRec’s prompts, domain adaptation, staged memory synthesis, ranking, propagation, rationale evaluation, and neighbor-curation comparisons. These materials emphasize compact collaborative contexts, asynchronous updates, and interpretable zero-shot curation.

  • F Prompt Templates and Contexts: The appendix provides complete prompt templates for LMMem memory management and LLMRec reasoning to support reproducibility.
  • F.1 Meta-Prompt Template: Domain-specific metadata, statistics, interaction modes, and characteristics guide LMMem to generate interpretable curation rules offline for zero-shot neighbor pruning.
  • F.2 Domain-Specific Prompt Contexts: Table 8 reports per-stage input and output token consumption, highlighting high input/output ratios in memory stages under asymmetric commercial-LLM pricing.
  • F.3 Stage-R Memory Synthesis Prompt: The Stage-R synthesis prompt distills user and neighbor memories into structured, confidence-scored memory facets that form collaborative context for downstream reasoning.
  • F.4 Stage-ReRank Scoring Prompt: The Stage-ReRank prompt combines user instructions, candidate-item details, and collaborative memory facets to produce a 0-to-1 relevance score and rationale.
  • F.5 Stage-W Propagation Prompts: Following interaction, Stage-W asynchronously updates user and item memories while propagating insights to relevant neighboring nodes in one concurrent operation.
  • F.6 Rationale Quality Evaluation Protocol: GPT-4o independently scores rationales from three models on Specificity, Relevance, and Factuality using a 1–5 Likert scale at temperature 0.
  • G.1 Comparison of Curation Approaches: The neighbor-curation comparison presents LLM-generated rules as a zero-shot balance of domain adaptivity, interpretability, and low online inference cost, alongside quantitative filtering analysis.

G.2 Quantitative Analysis of LLM Curation

The quantitative analysis evaluates LLM-guided curation as a way to filter noisy collaborative neighbors and generate domain-specific rules. It describes the filtering proxy, example rules, and the prompts supporting synthesis, ranking, propagation, and rationale evaluation.

  • Quantitative filtering analysis: The filtering analysis uses a 1,000-user subset containing 15,782 neighbors and defines low-similarity users and outdated items as likely irrelevant neighbors.
  • Curation comparison: The generic-rule comparison applies equal weights to normalized recency, co-interaction count, and metadata-overlap features without domain-specific semantic adjustments.
  • Quantitative results: 73.8%: LLM curation reduced irrelevant item neighbors, serving as a noise buffer for the downstream reasoning agent.
  • Books domain rules: Books rules boost items with metadata_overlap_score > 0.6 by 2.5x because genre and author similarity are treated as strong signals in content-driven recommendations.
  • MovieTV domain rules: MovieTV rules apply stronger exponential recency decay after 60 and 180 days because preferences are characterized as volatile.
  • Stage-R synthesis: Stage-R uses personal and collaborative neighbor memories plus candidate items to synthesize preference facets for ranking.
  • Stage-ReRank and Stage-W: Stage-ReRank scores candidate-item fit using user requests, collaborative facets, and item memories, while Stage-W updates the user, clicked item, and relevant neighbors asynchronously.
  • Rationale evaluation: Rationale quality is evaluated by GPT-4o across three models and the Specificity, Relevance, and Factuality criteria using a standardized judge prompt.
Loading 2601.08816v3…