Source-linked AI summary
AtomRec: Evolving Atomic Memory for Agentic Recommendation
Peiyu Hu, Weihai Lu, Siying Gu, Zhuodong Liu, Zhaokai Luo, Yuean Niu, Zhiyong Wang, Jia Wang
TL;DR
Existing agentic recommender memories can lose fine-grained preference evidence by compressing it into coarse summaries and scalar links. ATOMREC uses evolving atomic memories with semantic links and evidence-path retrieval, and it consistently improves over baselines across four benchmarks. The paper also identifies limitations involving false positives, short-term exploratory interests, and post-hoc traceability.
Problem
Existing memory mechanisms compress evolving preference traces into coarse summaries and scalar collaborative links, limiting fine-grained and interpretable evidence for ranking.
Method
ATOMREC represents user and item memories as structured atomic notes, builds semantic links among them, and evolves related historical memories for context-aware evidence-path retrieval.
Results
ATOMREC consistently outperforms traditional, LLM-based, agentic, and memory-augmented baselines across four instruction-augmented benchmarks.
Takeaways & Limitations
Atomic construction, semantic collaborative linking, and dynamic evolution support fine-grained long-term preference modeling with evidence paths.
Takeaways & Limitations
Memory evolution can over-compress interests, under-rank short-term exploratory targets, and weaken post-hoc traceability when collaborative signals are absorbed into fields.
Abstract
from arXiv · showhide
Agentic recommender systems use large language models to maintain semantic memory and support evidence-aware recommendation. However, existing memory mechanisms often compress user and item information into coarse summaries and connect them with scalar collaborative links, making it difficult to preserve fine-grained preference stages or retrieve interpretable evidence as user interests evolve. We propose \textsc{AtomRec}, an agentic recommender with evolving atomic collaborative memory. \textsc{AtomRec} represents user and item memories as structured atomic units, builds semantic links across related memories, and evolves related historical fields when new interactions arrive. During recommendation, it retrieves linked memories as multi-hop evidence paths rather than isolated neighbor summaries, allowing collaborative signals to support grounded ranking. Experiments on four public benchmarks show that \textsc{AtomRec} consistently outperforms state-of-the-art agentic and memory-augmented baselines, with around 8.5\% average relative improvement across metrics.
1 Introduction
Agentic recommenders need memory that preserves evolving, fine-grained preference evidence and connects collaborative signals to interpretable ranking. ATOMREC addresses these limitations with linked atomic memories, dynamic evolution, and evidence-path retrieval, outperforming strong baselines across four benchmarks.
- Motivation: Existing agentic recommender memories compress evolving preference traces into coarse summaries and often retrieve isolated neighbor summaries instead of interpretable evidence paths.These limitations can weaken fine-grained evidence for ranking as user interests change.
- Approach: ATOMREC represents user and item memories as linked, evolving atomic units with structured semantic fields for locating, revising, and preserving specific preference traces.Atomic units serve as operational objects for memory reorganization rather than passive summary accumulation.
- Approach: Semantic Collaborative Link Construction builds relations across user and item atoms, turning scalar collaborative signals into interpretable evidence.The linking mechanism connects related memories so collaborative context can support evidence-aware recommendation.
- Approach: Dynamic memory evolution updates related historical atoms and links when new interactions provide additional evidence, enabling the memory graph to reorganize over time.This supports context-aware retrieval as preferences evolve.
- Results: ATOMREC outperforms strong baselines across four public benchmarks with around 8.5% average relative improvement and better handling of stronger preference drift.The reported evaluation covers Books, Goodreads, MovieTV, and Yelp.
2 Method
ATOMREC builds recommendation memory from structured atomic notes connected by semantic links, then evolves related historical fields and retrieves linked evidence paths for context-aware ranking.
- Atomic Collaborative Memory Construction: ATOMREC represents user and item memories as structured atomic notes rather than coarse summaries, enabling fine-grained retrieval and revision.Each note exposes complementary semantic fields, including content, keywords, tags, context, and links.
- Semantic Collaborative Link Construction: Semantic Collaborative Link Construction converts nearby atomic memories into interpretable relations such as shared topics, intent progression, and preference transitions.The linking agent retrieves semantically nearby memories, analyzes them jointly, and stores linked identifiers with optional relation descriptions.
- Dynamic Memory Evolution: Dynamic memory evolution updates related historical notes when new evidence arrives, revising semantic fields and links while preserving each note’s original timestamp.The trigger uses the new note’s semantic links, similarity score, and an evolution threshold; updates can strengthen keywords, refine tags, and reconstruct context.
- Context-Aware Collaborative Retrieval and Recommendation: Context-aware retrieval expands relevant memories through semantic links to construct multi-hop subgraphs containing evidence from related users and items.Relation descriptions support synthesis, while hop depth and retrieval similarity control expansion.
- Context-Aware Collaborative Retrieval and Recommendation: The recommendation agent scores candidates using user memory, synthesized collaborative evidence, and item memory, receiving evidence paths instead of disconnected neighbor summaries.This combines semantic reasoning, collaborative enhancement, and memory evolution in a unified recommendation process.
3 Experimental Setup
The experiments evaluate ATOMREC against traditional, language-model, agentic, and memory-augmented recommenders using full test sets and controlled ablation protocols. They also examine component contributions, preference drift, qualitative interpretability, and hyperparameter sensitivity.
- Research Questions: Experiments address overall accuracy, component effects, preference drift, and qualitative behavior of evolving atomic memory.The component analysis includes atomic construction, semantic linking, memory evolution, link-aware retrieval, and key hyperparameters.
- Baselines: ATOMREC is compared with traditional, LM-based, agentic, and memory-augmented baselines, including LightGCN, SASRec, P5, Vanilla LLM, iAgent, RecBot, AgentCF, i2Agent, and MemRec.The baselines span three comparison categories specified in the evaluation protocol.
- Evaluation Protocols: Main results use full test sets with candidate size N=10 and report H@K and N@K for K∈{3, 5}.Ablation and analysis experiments use the same randomly sampled 1,000 users across methods.
- Implementation Details: The main experiments use gpt-4o-mini for memory and recommendation agents and frozen Sentence-T5 for atomic-field embeddings.Unless otherwise specified, k_link=20, τ_evo=0.7, and h=2 are selected on validation data and examined in sensitivity analysis.
4 Experimental Results
ATOMREC achieves the best results across four datasets and all metrics, with approximately 8.5% improvement over the strongest baseline. Ablations, sensitivity analysis, preference-drift evaluation, and case studies support the value of atomic memory, semantic links, evolution, and path-aware retrieval.
- Overall Performance: ATOMREC achieves the best results on all four datasets and all metrics, improving over the strongest baseline by around 8.5%.The reported comparison covers Books, Goodreads, MovieTV, and Yelp.
- Ablation Study: Removing any component reduces performance, indicating that atomic memory, semantic linking, memory evolution, and link-aware retrieval work jointly.The ablation study is conducted on Books, with additional datasets reported in the appendix.
- Ablation Study: Collapsing atomic notes into coarse summaries causes the largest ablation drop, while embedding-only neighbors and removed links also reduce performance.These results support structured fields and relation-aware evidence over coarse or embedding-only alternatives.
- Hyperparameter Sensitivity: ATOMREC performs best on Books with 𝑘link = 20, 𝜏evo = 0.7, ℎ = 2, and 𝑘ret = 15.The sensitivity analysis suggests moderate linking, conservative evolution, limited multi-hop retrieval, and a compact candidate set.
- Preference Drift Analysis: All methods degrade as preference drift increases, but ATOMREC’s gains over MemRec become larger for higher-drift users.The analysis divides users into Low, Medium, and High Drift groups using early- and late-history segments.
- Case Study: In a representative case, full ATOMREC follows 1-hop and 2-hop evidence links to rank the target item first after the user’s interests shift across stages.Without atomic memory the target is missed, while link-aware retrieval without atomic memory provides only a partial match.
- Efficiency: ATOMREC achieves higher N@5 than representative LLM-based and agentic baselines with moderate extra cost, while using more tokens than MemRec.The reported overhead comes from atomic construction, semantic linking, field-level evolution, and evidence synthesis.
5 Related Work
Related work spans external memory for LLM agents and agentic recommendation systems. Existing approaches include workflow-based memory, associative linking, simulation, planning, tool use, and memory-driven personalization, while collaborative memory across users and items remains underexplored.
- Memory in LLM Agents: LLM-agent memory systems evolved from predefined workflows and hierarchical buffers toward more adaptive associative linking and layered memory.Examples include MemoryBank, MemGPT, SCM, A-Mem, Mem0, and MemInsight.
- Memory in LLM Agents: Existing LLM-agent memory work predominantly focuses on single-agent memory, leaving collaborative memory across users and items underexplored for interactive recommendation.This identifies the collaborative-memory setting addressed by ATOMREC.
- LLM-based Recommendation Agents: LLM-based recommendation agents use simulation, planning, tool use, and memory-driven personalization to support recommendation workflows.Representative systems include Agent4Rec, AgentCF, RecMind, InteRecAgent, MACRec, ChainRec, RecThinker, and RecNet.
6 Conclusion
ATOMREC combines structured atomic memories, semantic links, and evolving historical fields for agentic recommendation. Across four benchmarks, it improves over diverse baselines, while ablations and qualitative analysis support its component design and evidence-path retrieval.
- Conclusion: ATOMREC represents user and item memories as structured atomic notes, connects them through semantic links, and evolves related historical memories as interactions arrive.This replaces coarse summaries and scalar collaborative edges with atomic units and semantic evidence paths.
- Conclusion: Experiments on four instruction-augmented benchmarks show consistent improvements over traditional, LLM-based, agentic, and memory-augmented baselines.Ablation studies and qualitative analysis support atomic construction, semantic collaborative links, and dynamic evolution.
- Conclusion: The results identify more structured, linkable, and evolvable memory as a promising direction for interpretable and adaptive recommender agents.
Limitations
ATOMREC has three limitations involving fine-grained discrimination, exploratory preferences, and provenance preservation. These limitations define important boundaries for current use and motivate future work.
- Limitations: Memory evolution may over-compress interests into a dominant subtheme, causing semantically adjacent false positives when fine-grained candidate discrimination is required.The model may identify the correct broad preference region but under-rank the target item.
- Limitations: ATOMREC may overemphasize stable long-term preference trajectories and under-rank idiosyncratic or short-term exploratory targets.
- Limitations: Collaborative signals may be absorbed into memory fields without explicit final-step links, weakening post-hoc traceability.The paper proposes provenance preservation for semantic memory links as future work.
Coarse Memory Atomic Memory
The paper contrasts coarse memory rewriting with ATOMREC’s atomic memory approach for preserving evolving preference evidence and supporting linked retrieval. The supplied case study shows atomic memory recovering a target item that coarse summaries miss or only partially match.
- Coarse Memory: Coarse summaries repeatedly rewrite broad preference profiles, weakening stage-level evidence as interests evolve.The case study traces shifts across spiritual communication, grief healing, and afterlife-related interests.
- Atomic Memory: ATOMREC retrieves the target item after coarse memory misses it and a partial-match approach fails to fully connect the relevant preferences.The target item is On the Edge of the Etheric.
- Atomic Memory: Link-aware retrieval connects preference evidence across memories instead of relying on isolated or broad summaries.The case study describes multi-hop links connecting grief healing with afterlife evidence and etheric explanations.
C Implementation Details
ATOMREC constructs atomic user and item memories, links semantically related notes, evolves historical fields, and retrieves linked evidence for ranking. Its implementation uses deterministic agents, temporal controls, and tuned memory-side hyperparameters.
- Memory Construction: Each interaction creates at most one user atomic note and one item atomic note, stored in a shared memory space.Atomic notes include textual content, timestamps, keywords, semantic tags, contextual descriptions, embeddings, and linked identifiers.
- Semantic Linking: Semantic Collaborative Link Construction links each new note to nearby memories using similarity retrieval followed by structured semantic relation generation.The default linking size is k_link = 20.
- Memory Evolution: Dynamic memory evolution updates related historical fields and recomputes embeddings while preserving the original timestamp.Invalid or unsupported updates leave the previous note unchanged.
- Retrieval and Ranking: During recommendation, retrieved notes expand through semantic links up to hop depth h=2 before evidence synthesis and candidate ranking.Relation descriptions are used as textual evidence rather than learned edge weights.
- Evaluation Protocol: The evaluation uses pre-target interactions only, while default settings include k_link = 20 and τ_evo = 0.7.The held-out target is exposed only during final reranking, and malformed structured outputs receive one retry.
E Additional Ablation Results
Ablations and controls attribute ATOMREC’s gains to atomic memory structure, semantic links, memory evolution, and link-aware retrieval rather than simply expanded context. The temporal evaluation protocol excludes future interactions from memory construction and evolution.
- Ablation Results: Removing the full atomic memory representation causes the largest performance drop on Books and Goodreads.Field-specific ablations also show contributions from keywords, tags, and contextual descriptions.
- Ablation Results: Embedding-only links, append-only memory, and link-aware retrieval removal all underperform the full model.These results support relation-aware linking, field-level evolution, and connected evidence-path retrieval.
- Cost-Matched Control: 6.5%: Cost-matched ATOMREC still outperforms MemRec on average under a similar token budget.The control retains atomic construction, semantic linking, and field-level evolution while limiting retrieved notes and evidence tokens.
- Evolution Audit: Evolution audits find that most updated notes preserve original semantics and are supported by nearby evidence, although not every update improves ranking.Harmful cases mainly involve over-merging adjacent themes or absorbing short-term exploration into dominant long-term preferences.
- Leakage Control: The temporal protocol constructs, links, evolves, and retrieves memory using only interactions before the target time.The target item appears only as a candidate during final reranking and is not written back for that evaluation instance.
H Semantic Link Quality Analysis
The analysis evaluates whether ATOMREC’s semantic links provide useful relational evidence and whether its atomic memory remains manageable as interactions accumulate. Results support higher-quality links, moderate graph growth, and limited online overhead.
- Semantic Link Quality: ATOMREC produces higher-quality semantic links than embedding-only linking across relatedness, relation correctness, and recommendation usefulness.The sampled links are rated on a 1–5 scale.
- Semantic Link Quality: LLM-guided linking filters weakly useful neighbors and identifies why linked notes relate and whether they support downstream ranking.
- Memory Growth: The memory graph remains moderately sized, with fewer than three average links per note under the default setting.Evolution updates a limited set of related historical notes rather than repeatedly rewriting the entire memory space.
- Efficiency: ATOMREC uses more tokens than MemRec because of atomic construction, semantic linking, and field-level evolution, but the increase remains moderate.These memory-side operations can be cached and executed asynchronously.
- Efficiency: The default gpt-4o-mini backbone provides a favorable cost-performance trade-off because stronger closed backbones yield only modest additional Books N@5 gains at much higher relative cost.Reported costs are deployment comparisons rather than universal prices.
L Additional Qualitative Analysis
The qualitative cases show that evolving atomic memory can support accurate recommendations when memory aligns with the target, but errors arise from over-narrow compression, short-term deviations, and weakened collaborative-link traceability.
- Over-narrowed memory: Over-narrowed memory captures the correct broad preference region but causes semantically adjacent false positives through insufficient fine-grained discrimination.The error occurs among relevant candidates rather than wholly irrelevant recommendations.
- Idiosyncratic target deviation: A short-term or idiosyncratic target deviation can be under-ranked when the evolved memory coherently follows the user’s dominant long-term trajectory.In the example, The King of Style: Dressing Michael Jackson ranked third behind two other items.
- Successful alignment: Atomic memory supports accurate top-ranked recommendation when the target aligns with the evolved memory center.The successful case contrasts with cases involving overly narrow subthemes or sharp deviations from dominant preferences.
- Collaborative-link fidelity: Collaborative signals may shape memory semantics without remaining explicit and traceable as final-step links.The cases motivate preserving link persistence, confidence-aware decay, and provenance during memory evolution.
- Deployment considerations: Deployment raises privacy, governance, latency, and backbone-dependence concerns beyond the qualitative recommendation cases.The paper notes sensitive preference traces, incomplete end-to-end latency analysis, and evaluation centered primarily on gpt-4o-mini.