PaperlayerView exact revision on arXiv

Source-linked AI summary

Long-Term Agent Memory through Complementary Curation and Consolidation

Abhishek Gupta, Dongge Han, Daniel Madrigal Diaz, Wei-Ning Chen, Mayukh Das, Victor Rühle, Saravan Rajmohan

arXiv:2610.39790v1cs.AIcs.MA

TL;DR

Long-horizon dialogue agents need to recall facts, events, and relationships across interaction histories. The paper combines episodic and structured relational memory with offline consolidation; the hybrid system outperforms baselines on both benchmarks, reaching 90.2% on LoCoMo.

  • Problem

    Long-horizon agents remain limited in recalling, integrating, and reasoning over interaction histories.

    Source: §1 · p. 1

  • Method

    The system stores timestamped episodic memories and contextualized relational tuples, then clusters, merges, or deletes entries during offline consolidation.

    Source: §1 · p. 1, §3.2 · p. 3

  • Results

    The hybrid system outperformed state-of-the-art baseline methods on both benchmarks, reaching 90.2% on LoCoMo.

    Source: §1 · p. 1, §4.2 · p. 4

  • Takeaways & Limitations

    The study concludes that combining both memory types performs strongly across the two benchmarks, while consolidation improves efficiency without much performance loss.

    Source: §6 · p. 5

    Evaluation covers textual conversations, and changing the backbone from gpt-4.1-mini to gpt-4o-mini corresponded to a 6.7 percentage point drop.

    Source: §7 · pp. 5–6

Abstract

from arXiv · show

Long-horizon dialogue agents require effective memory systems to recall facts, events, and relationships from interaction histories. In this work, we propose a hybrid-memory architecture comprising episodic memories and structured relational memories and a consolidation mechanism which preserves efficiency. We conduct an empirical study across two benchmarks: LoCoMo and LongMemEval-S. Our experiments demonstrate that the two memory types are consistently complementary, combining both outperform state-of-the-art memory frameworks and achieves 90.2% accuracy on LoCoMo and 85.5% on LongMemEval-S.

1 Introduction

Long-horizon agents struggle to recall, integrate, and reason over interaction histories, motivating memory systems that selectively retrieve past experience. The paper proposes complementary episodic and relational stores with offline consolidation and reports strong results on two dialogue benchmarks.

1 Introduction

  • Long-term memory systems selectively retrieve historical experience, mitigating the need for massive context windows and fostering more coherent, efficient agent behavior.Source: §1 · p. 1
  • Episodic memory preserves event sequence and context, while structured relational memory makes entity-specific connections explicit for cross-session inference.The system also consolidates related or redundant entries and removes conflicting or outdated ones.Source: §1 · p. 1
  • The hybrid system outperformed state-of-the-art baseline methods on both LoCoMo and LongMemEval-S, while consolidation compressed memory by 16.34% without much performance loss.On LoCoMo, consolidation also improved accuracy on both single-hop and multi-hop QA tasks.Source: §1 · p. 1

2 Related Work

Prior long-term memory work includes context-window management and systems that extract, organize, or route information across sessions. This paper combines episodic memory with relational tuples and offline consolidation.

2 Related Work

  • Prior systems manage context between windows and external storage, maintain salient content across sessions, build temporal graphs, or distill coherent episodes.Source: §2 · pp. 1–2
  • The system combines episodic memory with a lightweight relational tuple store grounded by timestamps and source context, then consolidates redundant memories offline.The authors contrast this with prior approaches that manage context or maintain salient content across sessions.Source: §2 · pp. 1–2

3 Method

The method curates episodic and relational memories from each conversation session, consolidates similar entries offline, and retrieves from both stores to answer queries.

3 Method

  • The system has an offline session-level curation phase, an offline consolidation phase, and a retrieval-and-generation phase for user queries.Source: §3 · p. 2

3.1 Memory Curation

  • For each session, episodic records preserve a title, summary, narrative details, and timestamp, while relational tuples encode entities, relations, objects, timestamps, and source context.The contextual description records the circumstances in which a tuple was mentioned, and the timestamp grounds it temporally.Source: §3.1 · pp. 2–3

3.2 Memory Consolidation

  • Offline consolidation clusters semantically similar memories with HDBSCAN, then merges redundant or complementary entries and deletes outdated, conflicting, or superseded ones.This targets memory-store growth, storage costs, retrieval overhead, and noise.Source: §3.2 · p. 3

3.3 Memory Retrieval & Question Answering

  • At query time, independent semantic searches retrieve the top-k episodic memories and top-2k relational tuples, with k set to 10 because performance plateaus beyond that value.Episodes are indexed by their details field, while tuples are indexed as natural-language strings.Source: §3.3 · p. 3

4 Experiments

The experiments evaluate the hybrid memory system on LoCoMo and LongMemEval-S against memory-framework baselines. The reported results cover benchmark performance, memory efficiency, and reasoning-task gains.

4.1 Experiment Setup

  • The evaluation uses LoCoMo, with 1,540 QA pairs across 10 dialogues, and LongMemEval-S, with 500 questions in chat histories averaging 115K tokens per question; baselines include FULLCONTEXT, RAG-4096, LANGMEM, MEM0, ZEP, and NEMORI.Table 1 reports LoCoMo LLM-as-a-judge, F1, and BLEU scores; Tables 2 and 3 cover LongMemEval-S and LoCoMo performance and efficiency.Source: §4 · p. 3, §4.1 · p. 3, Table 1 · p. 4

4.2 Results

  • 90.2% score on LoCoMo with gpt-4.1-mini; with gpt-4o-mini, the system scores 83.5% and outperforms all baselines.The gpt-4.1-mini result is slightly below FullContext’s 91.3% while using 87% fewer tokens.Source: Table 1 · p. 4, §4.2 · p. 4
  • 85.5% overall score on LongMemEval-S, surpassing FULLCONTEXT and NEMORI across nearly all evaluation categories.Table 2 reports detailed performance by question type.Source: Table 2 · p. 4, §4.2 · p. 4
  • The largest gains occur in Multi-Hop and Temporal reasoning, where the interaction between episodic and structured relational memories yields +6.3 and +6.8 points, respectively, on gpt-4o-mini.Episodic memory preserves a timestamped event timeline, while relational memory captures dependencies among entities and facts.Source: §4.2 · p. 4
  • 16.34% memory-store compression with only 1% performance degradation; on LoCoMo, the approach also improves accuracy by 9.1 points over NEMORI at comparable token use.The results compare performance and efficiency on LoCoMo.Source: §4.2 · p. 4, Table 3 · p. 4

5 Analysis and Ablations

The analysis tests backbone generalization and sensitivity to retrieval and consolidation settings. Results show a retrieval-budget plateau and a cluster-size trade-off between compression and accuracy.

5.1 Generalization Across Models

  • gpt-4.1-mini performs comparably to gpt-5.4, while Qwen3-8B reaches 76.2% overall accuracy, showing performance across backbones of varying capacity.The analysis uses LoCoMo; Table 4 reports framework performance by backbone.Source: §5.1 · p. 5, Table 4 · p. 5

5.2 Effect of Retrieval Budget

  • At k=2, the system retains a 0.811 score; performance rises sharply through k=10, then plateaus, with less than 1 point gain from k=10 to k=20.The retrieval-budget analysis uses gpt-4.1-mini and varies k for both episodic and relational memories.Source: §5.2 · p. 5

5.3 Effect of Consolidation Cluster Size

  • For mcs values from 5 to 20, LLM Score ranges from 0.892–0.898, within 1 point of the 0.902 no-consolidation baseline; mcs=2 drops performance by 2.6 points.Moderate cluster sizes (mcs∈[5, 10]) yield 16–22% compression with negligible accuracy loss; as mcs increases, compression falls from 30.9% to 9.6%, while clusters and corresponding LLM consolidation calls decrease by up to 15×.Source: §5.3 · p. 5, Table 5 · p. 5

6 Conclusion

The paper presents a hybrid agent memory system that curates and retrieves episodic and structured relational memories, evaluated on LoCoMo and LongMemEval-S. Combining the memory types delivers strong performance, while consolidation improves efficiency without much performance loss.

  • The system curates and retrieves both episodic and structured relational memories for agents.Source: §6 · p. 5
  • The empirical study evaluates the hybrid system on LoCoMo and LongMemEval-S.Source: §6 · p. 5
  • Combining both memory types delivers strong performance, and memory consolidation improves efficiency without much performance loss.Source: §6 · p. 5

7 Limitations

The evaluation focuses exclusively on textual conversational data, and memory curation quality varies across backbone LLMs. Supporting non-text modalities and developing more robust behavior across model families remain future directions.

  • The current evaluation focuses exclusively on textual conversational data, leaving non-text modalities such as vision and audio for future work.Source: §7 · p. 5
  • A shift from gpt-4.1-mini to gpt-4o-mini resulted in a 6.7 percentage point drop, reflecting variation in memory curation quality across backbones.Source: §7 · p. 6
  • Robust and consistent memory mechanisms across diverse model families remain an open research problem.Source: §7 · p. 6

A Prompt used for Episodic Memory Curation

The appendix prompts specify how episodic memories and relational triplets are curated, including temporal precision, detail preservation, and structured output formats.

A Prompt used for Episodic Memory Curation

B Prompt used for Relational Memory Curation

C Implementation details

The consolidation pipeline reduces embedded memories before clustering them. It uses UMAP for dimensionality reduction and HDBSCAN for clustering, with specified distance measures and parameters.

  • Stored memories use pre-computed 1536-dimensional embeddings from OpenAI’s text-embedding-3-small model as input to the pipeline.Source: C Implementation details · p. 9
  • UMAP projects the embeddings using cosine distance, 10 neighbors, 5 target components, and minimum distance zero.Source: C Implementation details · p. 9
  • HDBSCAN clusters the reduced representations using Euclidean distance, Excess of Mass selection, minimum cluster size 10, and minimum samples 1.Source: C Implementation details · p. 9

D Memory Component Analysis

The LoCoMo ablation evaluates episodic-only, relational-only, hybrid, and consolidated memory variants. The hybrid’s gains over either single memory type are largest on Multi-Hop and Temporal tasks.

D Memory Component Analysis

  • The LoCoMo ablation evaluates episodic-only, relational-only, full hybrid, and consolidated variants; Table 6 reports results for gpt-4o-mini and gpt-4.1-mini.Source: D Memory Component Analysis · p. 9, Table 6 · p. 9
  • +6.3 and +2.5 points on Multi-Hop, and +6.8 and +3.4 points on Temporal, are the hybrid gains over the stronger single memory type with gpt-4o-mini and gpt-4.1-mini, respectively.The two memory types are complementary on these tasks.Source: D Memory Component Analysis · p. 9
  • With gpt-4.1-mini, the hybrid improves Single-Hop and Open-Domain by 3.6 and 1.0 points, respectively, over the best single component.Source: D Memory Component Analysis · p. 9
  • With gpt-4o-mini, the hybrid matches the best single component on Single-Hop and Open-Domain.Source: D Memory Component Analysis · p. 9

E Extended Comparison with Prior Work

The comparison with prior work distinguishes memory systems by representation, curation strategy, and operational cost. The paper contrasts its two flat stores and session-level curation with alternatives’ graph, hierarchical, or turnwise approaches.

Representation

  • Unlike Zep’s temporal knowledge graph, A-Mem’s linked Zettelkasten-style notes, and H-Mem’s hierarchical indexing, this system uses episodic records and semantically retrieved relational tuples in two flat stores.This sacrifices native graph traversal while avoiding graph-construction and hierarchy-maintenance overhead.Source: E Extended Comparison with Prior Work · p. 9

Curation strategy

  • Unlike Mem0’s turn-by-turn add/update/delete/no-op operations and Nemori’s prediction-error-based distillation, this method curates each session into episodic and relational entries, then consolidates offline by clustering.Consolidation merges redundant memories and removes outdated ones.Source: E Extended Comparison with Prior Work · p. 9

Operational cost

  • Session-level curation amortizes extraction across turns, fixed top-k retrieval avoids extra query-time LLM calls, and offline consolidation can run every 10–20 sessions while preserving QA accuracy.Turnwise systems such as Mem0 and A-Mem scale LLM calls linearly with conversation length.Source: E Extended Comparison with Prior Work · p. 9
Loading 2610.39790v1…