acceptodds
Under review as a conference paper at ICLR 2027

CoRelMem: Coherent Relation Induction for Long-Term Agent Memory

Abstract

Long-term memory is essential for LLM-based agents to accumulate and reuse information across extended interactions. Existing memory systems commonly retrieve historical information based on lexical or semantic similarity, implicitly assuming that memories more similar to the query are also more useful for answering it. However, semantic similarity does not necessarily reflect reasoning relevance. Relevant memories with weak query similarity may be overlooked, although their relevance may be revealed through associations with other retrieved memories. Meanwhile, highly similar but low-value memories may receive excessive priority and occupy the limited retrieval budget. To address this issue, we propose CoRelMem, a structured coherent relation induction framework for long-term agent memory. CoRelMem organizes long-term memories through logical association and temporal evolution, enabling retrieval to follow cross-memory dependencies and recover relevant memories that cannot be identified through direct semantic matching alone. Based on these structures, a knowledge-guided dual-path retrieval strategy expands candidate memories from complementary relational and temporal perspectives, improving the coverage of useful information. CoRelMem further introduces an information-aware reranking mechanism that considers both retrieval relevance and conditional information contribution, reducing the priority of superficially similar but less informative memories. Experiments on LoCoMo and LongMemEval demonstrate consistent improvements in both memory retrieval and downstream question answering. The code is available at https://anonymous.4open.science/r/CoRelMem-DC8B.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Agent Memory, Knowledge Graph, Memory Retrieval

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_s44DOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

CoRelMem extracts atomic memories, builds an entity knowledge graph, and organizes memories into an association graph and temporal trajectories. At retrieval time, it expands initial candidates through both structures, then reranks evidence using semantic similarity and language-model conditional likelihood gain. The motivation is that useful evidence can have low query similarity, while highly similar memories can be uninformative. The main question is whether the proposed structures provide benefits beyond a stronger reranking pipeline with comparable context and computation.

Critical strengths
  1. The method provides a consistent improvement overall QA gains. Against LightMem, LoCoMo accuracy improves from 67.53 to 72.86 with Qwen and from 67.86 to 72.34 with Llama. LongMemEval-S improves from 46.20 to 49.20.
  2. Some results align with the motivation. Qwen’s LoCoMo multi-hop accuracy increases from 69.86 to 79.43, which supports investigating cross-memory retrieval. The paper includes ablations, retrieval metrics, parameter sweeps, and cost comparisons. The appendix also explains evidence bundles and disjoint selection more clearly than the main text.
Critical weaknesses
  1. The novelty of this work relative to existing methods is not clearly established. The comparison should cover a broader range of structured memory approaches, particularly methods for temporal graph construction and learning.

  2. The paper provides limited discussion of how CoRelMem differs from the selected baselines in its mechanisms and design choices. A more detailed comparison would help clarify its technical contributions. The evaluation is limited to long-term conversational memory benchmarks. The limitations of the proposed associative and temporal mechanisms, including the conditions under which they may fail, are also insufficiently discussed.

Questions for authors
  1. Could the authors compare CoRelMem with additional structured memory baselines, particularly temporal graph methods, and clarify how it differs from approaches that model similar dependencies?

  2. Beyond the reported benchmark improvements, could the authors explain the specific advantages of their design and provide evidence supporting the individual design choices?

  3. Could the authors evaluate CoRelMem on additional benchmarks and provide a failure-case analysis? Experiments with stronger LLM backbones would also help assess whether the gains persist across model capabilities and clarify the trade-off between performance and computational cost.

  4. Does the reported token cost for memory organization include all operations involved in coherent relation induction, including LLM-based relation judgments and conditional likelihood scoring?

5 comments
AuthorsOpenReview ↗

Response to Reviewer s44D (Part 1)

Thank you for the valuable comments. Below, we have carefully addressed each of your comments.

W1, 2.1&Q1: Novelty and comparisons with structured memory methods

We thank the reviewer for highlighting the need to clarify CoRelMem's methodological distinctions from existing structured-memory approaches. CoRelMem addresses two challenges that are not fully resolved by existing methods. Relevant memories may be overlooked due to weak semantic similarity to the query, while highly similar but low-value memories may be overprioritized. As illustrated in Figure 1(a)–(b), semantic similarity does not necessarily reflect the actual value of a memory for answering the current query. The key novelty of CoRelMem lies in expanding potentially relevant memories through cross-memory dependencies and prioritizing retrieved evidence based on its information contribution rather than semantic similarity alone.

The theoretical motivation for CoRelMem's dependency modeling is discussed in Appendix A.2. Inspired by Markov Random Fields and Markov chains, CoRelMem models logical associations through local conditional dependencies and temporal trajectories through direct state transitions within evolving event threads. These are modeling interpretations rather than explicitly learned probabilistic graphical models. In comparison, GAM constructs hierarchical memory graphs based on event progression and topic associations, while TSM organizes temporally related memories through semantic timelines and durative memories to preserve temporal continuity and validity. ActMem constructs semantic and causal memory connections, using embedding similarity to identify semantic relations and PMI-style likelihood differences to validate LLM-generated causal relations. Unlike these approaches, CoRelMem focuses on information dependencies between atomic memories and direct state transitions within evolving events, rather than relying primarily on structural associations or temporal continuity. This design enables retrieval to expand along memories that provide complementary information or represent successive states of the same event, helping recover relevant evidence that may be missed by direct semantic matching. Furthermore, IAR evaluates how much information each expanded candidate contributes to the current query, allowing CoRelMem to prioritize useful evidence while reducing interference from highly similar but irrelevant memories.

To provide stronger empirical evidence, we additionally evaluated GAM and TSM on LoCoMo using Qwen3-4B and GLM-5.3-Flash. CoRelMem consistently outperforms both baselines, with particularly notable improvements on multi-hop and temporal reasoning tasks. Detailed results are provided in our response to Reviewer nhct, W2. In the revised manuscript, we have incorporated the additional GAM and TSM results into Table 1 and further clarified the methodological distinctions between CoRelMem and existing structured-memory approaches in Section 2 and Appendices A.1.1 and A.1.2.

W2.2 & Q3.1: Additional Benchmark Evaluation

To address the concern regarding limited benchmark coverage, we conducted additional experiments on Long-MT-Bench+ using GLM-5.3-Flash, comparing CoRelMem against six baselines. CoRelMem achieves the best performance on five out of six metrics. The complete results are provided in our response to Reviewer nhct, W1. In the revised manuscript, we have added Appendix A.6, which presents the complete experimental results and corresponding analysis on Long-MT-Bench+.

AuthorsOpenReview ↗

Response to Reviewer s44D (Part 2)

W2.3 & Q3.1: Failure-Case analysis

We further conducted a failure-case analysis to investigate the limitations of association-based and temporal evidence expansion. As shown in the table, the full CoRelMem retrieves several related memories about John's charity activities and basketball career but fails to retain the specific event evidence in its final Top-10 results. In contrast, removing association-based expansion while retaining temporal retrieval allows the gold evidence to appear at rank 9, whereas removing both expansion paths also fails to retrieve it. This case illustrates a potential limitation of multi-path retrieval, as association-based expansion may introduce competing candidates that prevent relevant evidence from being retained in the final retrieval results. Although the two retrieval paths generally improve evidence coverage, their combination can occasionally introduce interference during candidate selection and prioritization. In the revised manuscript, we have expanded the discussion in Section 4.5 and added Appendix A.9 to provide a more detailed examination of this failure case and the limitations of multi-path evidence retrieval.

Question: What charity event did John organize recently in 2024?
Reference: benefit basketball game
Retrieval Setting: CoRelMem
#1 John is teaming up with a local organization that helps disadvantaged kids through sports and school, hoping to use his platform for positive impact on the community and to inspire others.
#2 John wants to use his platform to make a positive difference and inspire others, possibly by starting a foundation and doing charity work.
...
#8 Much happened with John off court in the last month.
#9 [Association] John used contacts in the basketball industry and his marketing skills to make connections for endorsements, with networking playing a big role (image shows a basketball card of a man holding a basketball).
#10 [Association] John feels fulfilled having the trust and admiration of young athletes who see him as a role model, and is glad his experiences help shape their future and inspire them to pursue their dreams.
Retrieval Setting: w/o Association Expansion
...
#9 [Trajectory] John held a benefit basketball game last week (shared a photo of a boy dribbling a basketball on a court). Gold
#10 John attended a Harry Potter trivia charity event last August.
Retrieval Setting: w/o Association & Temporal Expansion
...
#9 John attended a Harry Potter trivia charity event last August.
#10 Tim supports John's endorsement journey.

Q2: Specific advantages of the design

The advantages of CoRelMem can be explained from both theoretical and empirical perspectives.

Theoretically, CoRelMem integrates structured evidence expansion with information-aware prioritization to address the limitations of similarity-based memory retrieval. Specifically, the association graph enables the retrieval of relevant memories that are weakly matched to the query but connected through cross-memory dependencies. Temporal trajectories capture direct state transitions rather than merely chronological proximity, helping recover evidence distributed across evolving events. Furthermore, IAR mitigates the noise introduced by candidate expansion by prioritizing evidence based on both semantic relevance and conditional information contribution. Together, these designs improve evidence coverage while reducing the influence of less informative candidates. Their theoretical motivations are further discussed in Appendix A.2.

Empirically, the original ablations in Table 4 show that removing KG, CRI, or IAR generally reduces performance. We further conducted additional ablations comparing semantic-only, association-only, temporal-only, and dual-path retrieval, each with and without IAR. The results show that dual-path retrieval achieves the highest overall ACC among these configurations, while IAR consistently improves performance across all four retrieval strategies. Detailed results are provided in our response to Reviewer cSdp, W3 & Q2. In the revised manuscript, we have expanded the ablation analysis in Section 4.3 and added detailed comparisons of retrieval strategies with and without IAR in Appendix A.7.1.

AuthorsOpenReview ↗

Response to Reviewer s44D (Part 3)

Q3.2: Evaluation with Stronger LLM Backbones

To assess whether CoRelMem's improvements persist with a stronger LLM backbone, we conducted additional experiments using GLM-5.3-Flash on both LoCoMo and Long-MT-Bench+. On LoCoMo, CoRelMem achieves the highest overall ACC and F1, with particularly notable improvements on multi-hop and temporal reasoning tasks. On Long-MT-Bench+, CoRelMem also outperforms the baselines on five out of six evaluation metrics, as detailed in our response to Reviewer nhct, W1. Together with the original experiments using Qwen3-4B and Llama-3.2-3B, these results further demonstrate the effectiveness of CoRelMem across different LLM backbones. In the revised manuscript, we have added Appendix A.5 and Appendix A.6 to present the complete experimental results and corresponding analyses.

Table: QA Performance on LoCoMo Using GLM-5.3-Flash.

MethodACCF1B-1R-LMultiTempOpenSingle
Mem048.4417.6913.8416.8147.1646.7246.8749.7
MemGAS59.6816.6510.715.6741.1350.7851.0470.27
A-Mem50.9718.0714.0617.2547.8747.3537.554.93
LightMem72.8235.9317.1734.4166.6775.3145.8377.02
TSM58.4427.214.3626.4353.954.2140.6263.61
GAM*59.3524.9819.7424.0459.2258.5742.7161.59
CoRelMem78.1839.2318.1837.3174.4779.1346.8782.64

GAM: Reimplemented by us, as no official implementation is publicly available.

Q4: Scope of the reported token cost

The reported token cost in Section 4.6 and Figure 4 includes LLM-based relation judgments during association graph and temporal trajectory construction. CLG-based conditional likelihood scoring is not included, as it is performed through likelihood evaluation rather than text generation and incurs additional computational overhead. In the revised manuscript, we have clarified in Section 4.6 that all reported costs are measured during memory construction, with token consumption referring specifically to the output tokens generated by LLMs during this process.

Reviewer_cSdPOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes CoRelMem, which organizes long-term memories into an association graph and temporal evolution trajectories. It expands candidate evidence through two retrieval paths, then combines semantic similarity and conditional likelihood gain for filtering and reranking. The authors report improvements over the selected baselines on LoCoMo and LongMemEval. The overall method is complete, but the experiments do not sufficiently establish its distinct contribution over existing structured memory methods or show why the components need to be combined.

Critical strengths
  1. The motivation is clear. The paper addresses the gap between semantic similarity and the actual usefulness of evidence. It also studies how useful evidence can be missed when it has low similarity to the query.

  2. The method provides a complete pipeline for memory organization and retrieval. It reports overall improvements over the selected baselines on LoCoMo and LongMemEval and includes initial component ablations.

Critical weaknesses
  1. Limited methodological novelty. Associative memory, temporal structures, and multi-signal ranking have all been explored in prior work. ActMem also uses PMI-style likelihood differences to verify memory relations. This paper mainly combines these ideas and adds information-aware reranking. However, it does not clearly establish what is new beyond this combination.

  2. Missing direct comparisons with the closest prior work. The authors discuss GAM, TSM, and ActMem but do not include them in the main experiments. The current results therefore do not establish the added value over these closely related methods.

  3. Insufficient ablations and budget controls. The ablations do not clearly separate the gains from the association path, the temporal path, and their combination. They also do not isolate the contributions of candidate expansion and reranking. Moreover, one evidence unit can contain several memories. The same Top-K may therefore allow different context budgets, making the source of the gains unclear.

Questions for authors
  1. What is new beyond directly combining existing mechanisms from GAM, TSM, and ActMem? Could the authors add direct comparisons or controlled baselines within a shared framework?

  2. Could the authors compare semantic retrieval only, association expansion only, temporal expansion only, and dual-path expansion, each with and without IAR? This would help test whether the two paths complement each other and isolate the contribution of reranking.

  3. Do the gains remain when the final context tokens, number of atomic memories, and compute budget are matched? With a fixed candidate set, how does IAR compare with semantic-only ranking, CLG-only ranking, and a standard reranker?

2 comments
AuthorsOpenReview ↗

Response to Reviewer cSdP (Part 1)

Thank you for the valuable comments. Below, we have carefully addressed each of your comments.

W1&Q1.1:Methodological novelty

CoRelMem addresses two interconnected challenges in long-term memory retrieval. Relevant memories may be overlooked due to weak semantic similarity to the query, while highly similar but irrelevant memories may receive excessive priority. As illustrated in Figure 1(a)–(b), semantic similarity does not necessarily reflect the actual relevance of a memory to the current query. Consequently, relying solely on semantic matching or existing memory organization structures may fail to ensure both comprehensive evidence discovery and effective evidence selection.

The core novelty of CoRelMem lies in the effective expansion of potentially relevant memories and information-aware memory selection. For memory expansion, CoRelMem employs Coherent Relation Induction (CRI) to identify logical dependencies and temporal state transitions between atomic memories, using these relationships to guide retrieval expansion and discover relevant memories that may be missed by direct semantic matching. For memory selection, Information-Aware Reranking (IAR) evaluates the conditional information contribution of expanded evidence to the current query using Conditional Likelihood Gain (CLG), thereby reducing the priority of highly similar but irrelevant memories. Existing structured-memory approaches improve memory organization and retrieval but still have limitations in addressing these challenges. For example, GAM and TSM primarily focus on optimizing memory organization and the effective utilization of temporal information, respectively, and may still overlook valuable memories with weak semantic similarity to the query. ActMem combines causal relation modeling with counterfactual reasoning to discover potentially relevant memories through query augmentation and graph expansion. However, it lacks an explicit mechanism for selecting expanded evidence, making it difficult to distinguish highly similar but irrelevant memories from evidence that genuinely supports answering the query.

Therefore, the novelty of CoRelMem lies in addressing evidence omission and misprioritization by leveraging cross-memory dependencies for evidence expansion and incorporating conditional information contribution into evidence prioritization. Specifically, CoRelMem not only recovers relevant memories with weak semantic matches through logical and temporal dependencies but also employs IAR to assess how much information the expanded candidates contribute to the current query, prioritizing evidence that genuinely supports the query while suppressing highly similar but irrelevant memories. The theoretical motivations for these designs are discussed in Appendix A.2, and their effectiveness is supported by the original experiments and the additional ablation studies presented in our response to W3 & Q2. In the revised manuscript, we have further clarified the methodological distinctions of CoRelMem from existing approaches in Section 2 and Appendices A.1.1 and A.1.2.

W2&Q1.2: Comparisons with closely related methods

To address the concern regarding insufficient comparisons with closely related methods, we conducted additional experiments with GAM and TSM on LoCoMo using Qwen3-4B and GLM-5.3-Flash. As shown in table below, CoRelMem outperforms both methods across most evaluation metrics. Notably, CoRelMem also achieves substantial improvements on multi-hop and temporal reasoning questions, providing further evidence of the benefits of modeling cross-memory associations and temporal evolution. These results strengthen the empirical evaluation of CoRelMem against closely related structured-memory approaches. In the revised manuscript, we have incorporated the additional GAM and TSM results into Table 1, expanded the methodological comparisons in Appendices A.1.1 and A.1.2, and provided detailed descriptions of GAM and TSM in Appendix A.4.3.

BackbonemethodACCF1B-1R-1R-2R-LMultiTempOpenSingle
Qwen3-4bTSM53.8314.128.7715.57.9714.6651.4244.8631.2560.64
GAM*53.9613.779.1613.776.7112.9657.836.1434.3861.71
CoRelMem72.8637.3622.9637.4118.8535.5279.4362.3153.1276.93
GLM-5.3-FlashTSM58.4427.214.3627.6615.4826.4353.954.2140.6263.61
GAM*59.3524.9819.7424.9813.8824.0459.2258.5742.7161.59
CoRelMem78.1839.2318.1839.0221.8337.3174.4779.1346.8782.64

GAM: Reimplemented by us, as no official implementation is publicly available.

AuthorsOpenReview ↗

Response to Reviewer cSdP (Part 2)

W3&Q2: Insufficient ablations of retrieval strategies and IAR

To disentangle the contributions of different retrieval paths and IAR, we conducted additional ablation experiments comparing semantic-only retrieval, association-only expansion, temporal-only expansion, and dual-path expansion, each with and without IAR, on LoCoMo using GLM-5.3-Flash. The results show that dual-path retrieval achieves higher ACC than any individual retrieval strategy, supporting the complementary contribution of association-based and temporal expansion. Association-only and temporal-only retrieval perform less effectively than semantic retrieval, indicating that structured expansion is more beneficial as a complement to semantic matching rather than as a standalone retrieval strategy. IAR consistently improves all four retrieval configurations, demonstrating the contribution of IAR beyond candidate expansion. In the revised manuscript, we have updated Section 4.3 and added Appendix A.7, which provides detailed comparisons of different retrieval strategies with and without IAR.

SettingMethodRecall@10MRR@10nDCG@10
w/o IARsemantic68.9751.3651.02
association61.3849.2647.57
temporal39.5432.4530.03
dual-path69.2551.4651.17
IARsemantic72.3759.4957.23
association69.8262.3858.51
temporal45.6541.8937.94
dual-path73.6960.7158.48
SettdingmethodACCF1B-1R-1R-2R-LMultiTempOpenSingle
w/o IARsemantic72.5936.3917.2536.1619.8534.5970.5674.7645.8375.5
association65.9731.5315.1731.6616.6130.1262.7672.2739.5867.65
temporal47.2720.6211.8720.5910.0719.6152.8366.6730.2139.95
dual-path74.4836.6417.1436.6920.0535.0871.6377.5744.7977.65
IARsemantic76.6738.3317.5938.3221.1836.3975.1776.6343.7580.95
association73.6233.4515.7233.4118.2131.6676.1577.8840.6274.91
temporal52.7323.3712.723.2611.4422.0856.7375.3832.2945.06
dual-path78.1839.2318.1839.0221.8337.3174.4779.1346.8782.64

W3&Q3: Insufficient ablations of reranking strategies and budget controls

We conducted additional experiments on LoCoMo using GLM-5.3-Flash, examining the effects of evidence representation and different reranking strategies.

To investigate whether evidence units containing multiple atomic memories provide an advantage through additional context, we compared retrieving Top-10 evidence units with Top-10 atomic facts, where each atomic fact represents a single memory. As shown in table below, the two settings achieve nearly identical performance, with negligible differences in metrics. These findings suggest that the performance is not primarily driven by grouping multiple memories into evidence units. In the revised manuscript, we have included this comparison in Appendix A.7.3.

SettingRecall@10MRR@10nDCG@10ACCMultiTempOpenSingle
Evidence Unit73.6960.7158.4878.1874.4779.1346.8782.64
Atomic Fact73.6660.5758.3878.0575.5380.0644.7981.92

As shown in Figure 3(b) of the manuscript, we have examined the effect of varying the fusion weight between semantic relevance and CLG, including the semantic-only and CLG-only settings. To further address the concern, we provide an comparison of these two strategies and IAR under the dual-path retrieval setting. As shown in table below, IAR achieves the best overall performance. These results further support the effectiveness of jointly considering semantic relevance and conditional information contribution in evidence prioritization. In the revised manuscript, we have included these additional comparisons in Appendix A.7.2.

SettingRecall@10MRR@10nDCG@10ACCMultiTempOpenSingle
semantic-only69.2551.4651.1774.4871.6377.5744.7977.65
CLG-only72.9160.6558.2577.0173.4076.6344.7982.04
IAR73.6960.7158.4878.1874.4779.1346.8782.64
Reviewer_nhctOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes CoRelMem, a long-term agent memory framework that organizes memories into an association graph and temporal trajectories to retrieve evidence beyond direct semantic matching. The proposed approach reports improvements on LoCoMo and LongMemEval-S.

Critical strengths
  • The paper studies an important problem in memory retrieval that aims to retrieve evidence beyond direct semantic matching.
  • The combination of associative and temporal structures is well motivated.
Critical weaknesses
  • Limited benchmark coverage. The evaluation is restricted to LoCoMo and LongMemEval-S, leaving it unclear whether the gains generalize beyond these two conversational QA benchmarks. Broader evaluation on additional memory benchmarks and long-context reasoning tasks, such as MemBench, MemoryBench, NarrativeQA, DetectiveQA, NoCha, and PRELUDE, would help establish that the improvements are not specific to the selected datasets.

  • Insufficient comparison with closely related methods. Even on the two evaluated datasets, the experiments do not establish state-of-the-art performance. Comparisons omit closely related structured-memory methods discussed in the paper, such as GAM and ActMem. Also there are a wide range of structured memory approaches proposed for long context understanding such as CAM. These approaches should be compared as well.

Questions for authors

Please refer to the concerns raised above. Evaluating the proposed method on diverse and rigorous memory benchmarks would help me better assess the significance and practical value of this work.

3 comments
AuthorsOpenReview ↗

Response to Reviewer nhct

Thank you for the valuable comments. Below, we have carefully addressed each of your comments.

W1: Limited benchmark coverage

We conducted additional experiments on Long-MT-Bench+ using GLM-5.3-Flash, extending our evaluation beyond the original LoCoMo and LongMemEval-S benchmarks, as detailed in Appendix A.6 of the revised manuscript. We also evaluated CoRelMem and the baselines on LoCoMo using GLM-5.3-Flash to further assess performance across different LLM backbones, with the results presented in Appendix A.5. As shown in the tables, CoRelMem achieves the best performance on five out of six metrics on Long-MT-Bench+. On LoCoMo, it also achieves the highest ACC and F1 under GLM-5.3-Flash, with notable improvements on multi-hop and temporal reasoning. Together with the original results on LoCoMo and LongMemEval-S, these additional findings show consistent performance gains across three long-term memory benchmarks and three LLM backbones (Qwen3-4B, Llama-3.2-3B, and GLM-5.3-Flash), providing further evidence that the effectiveness of CoRelMem is not limited to the original evaluation settings.

Table R1: Results on Long-MT-Bench+ using GLM-5.3-Flash.

MethodACCF1B-1R-1R-2R-L
Mem074.6526.3117.5130.6311.9320.81
MemGAS74.3125.2416.5129.1813.2720.55
A-Mem78.1227.316.8631.0114.9122.45
LightMem73.9527.5518.4331.7413.3622.39
TSM74.6527.7616.8131.4115.1622.81
GAM*71.8827.4816.8531.2315.2622.61
CoRelMem82.9828.5317.2532.0215.8623.41

Table R2: Results on LoCoMo using GLM-5.3-Flash.

MethodACCF1B-1R-LMultiTempOpenSingle
Mem048.4417.6913.8416.8147.1646.7246.8749.7
MemGAS59.6816.6510.715.6741.1350.7851.0470.27
A-Mem50.9718.0714.0617.2547.8747.3537.554.93
LightMem72.8235.9317.1734.4166.6775.3145.8377.02
TSM58.4427.214.3626.4353.954.2140.6263.61
GAM*59.3524.9819.7424.0459.2258.5742.7161.59
CoRelMem78.1839.2318.1837.3174.4779.1346.8782.64

GAM: Reimplemented by us, as no official implementation is publicly available.

W2: Insufficient comparison with closely related methods

We conducted additional experiments with GAM and TSM on LoCoMo using Qwen3-4B and GLM-5.3-Flash. As shown in table below, CoRelMem outperforms both methods across most evaluation metrics. These results further demonstrate the effectiveness of CoRelMem compared with closely related structured-memory methods, particularly on multi-hop and temporal reasoning tasks, where modeling cross-memory associations and temporal evolution is especially important. In the revised manuscript, we have also expanded Section 2 (Related Work) and Appendices A.1.1 and A.1.2 to clarify the methodological distinctions between CoRelMem and related approaches, including ActMem, CAM, GAM, and TSM.

Table R3: Additional Comparisons with Structured Memory Baselines on LoCoMo.

BackboneMethodACCF1B-1R-1R-2R-LMultiTempOpenSingle
Qwen3-4BTSM53.8314.128.7715.507.9714.6651.4244.8631.2560.64
GAM*53.9613.779.1613.776.7112.9657.8036.1434.3861.71
CoRelMem72.8637.3622.9637.4118.8535.5279.4362.3153.1276.93
GLM-5.3-FlashTSM58.4427.2014.3627.6615.4826.4353.9054.2140.6263.61
GAM*59.3524.9819.7424.9813.8824.0459.2258.5742.7161.59
CoRelMem78.1839.2318.1839.0221.8337.3174.4779.1346.8782.64

Thank you for the additional experiments. While the new results broaden the evaluation, they do not fully address my concern about the scope of the tasks. Long-MT-Bench+ is another benchmark centered on QA over conversational history, whereas my review specifically asked for evaluation beyond this setting.

More specifically, I believe the work would be stronger with evidence along either of the following directions:

  • Memory construction and usage for downstream tasks, e.g., tasks such as those listed in my review.
  • Shift the focus to long-context understanding instead of agent memory, specifically by adding narrative understanding results, which are also the settings where related work such as CAM has been evaluated.

Thank you for clarifying the concern regarding task diversity. To further address this concern, we conducted an additional preliminary evaluation on DetectiveQA, one of the benchmarks suggested in the original review. DetectiveQA is a long-context narrative reasoning benchmark based on detective novels, requiring the integration of clues distributed throughout narrative texts. We randomly selected 6 samples comprising 40 questions from the human_anno subset (sample IDs: 117, 126, 140, 150, 198, and 209) and evaluated CoRelMem against LightMem. As shown in the table, CoRelMem outperforms LightMem across all reported retrieval metrics and answer accuracy. These results provide preliminary evidence of CoRelMem's applicability to long-context narrative reasoning beyond conversational history QA.

Although this evaluation is preliminary in scale, it directly examines the narrative understanding setting. We plan to further expand the evaluation to more samples and additional baselines and incorporate the extended results into a subsequent manuscript revision.

MethodRecall@3Recall@5Recall@10ACC
CoRelMem304057.567.5
LightMem22.537.55562.5

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

We sincerely thank the three reviewers (nhct, cSdP, and s44D) for their careful reviews and recognition of the importance of our research problem and motivation. During the rebuttal period, we conducted additional experiments and revised the relevant sections of the manuscript to address the reviewers' concerns. We summarize the main concerns and our responses below:

  • Methodological Novelty and Distinctions from Existing Work (cSdP, s44D): s44D and cSdP raised concerns about CoRelMem's methodological novelty and distinctions from existing methods. In response, we clarified that CoRelMem's central contribution lies in jointly addressing memory omission and misprioritization through CRI and IAR. In addition, we expanded our discussion of the differences between CoRelMem and related methods. These clarifications have been incorporated into Section 2 and Appendices A.1.1 and A.1.2.
  • Additional Comparisons with Related Baselines (nhct, cSdP, s44D): All three reviewers noted the lack of experimental comparisons with closely related structured-memory methods. To address this concern, we added comparisons with GAM and TSM on LoCoMo using both Qwen3-4B and GLM-5.3-Flash. The additional results have been incorporated into Table 1 and Appendix A.5, with implementation details provided in Section 4.1 and Appendix A.4.3.
  • Additional Benchmark Evaluations (nhct, s44D): nhct and s44D suggested evaluating CoRelMem on broader benchmarks. In response, we added Long-MT-Bench+ experiments with six baselines. Furthermore, in the follow-up discussion, nhct further noted that evaluation beyond conversational history QA, such as long-context narrative understanding, would strengthen the work. Accordingly, we conducted a preliminary evaluation on DetectiveQA, providing initial evidence of CoRelMem's applicability to narrative reasoning. We plan to expand this evaluation with more samples and baselines in a subsequent revision. Both evaluations have been incorporated into Appendix A.6.
  • Additional LLM Backbone Evaluations (s44D): s44D suggested evaluating CoRelMem with a stronger LLM. In response, we introduced GLM-5.3-Flash as an additional LLM backbone and conducted experiments on LoCoMo and Long-MT-Bench+. The additional results further support the effectiveness of CoRelMem across different LLM backbones. These experiments and analyses have been included in Appendices A.5 and A.6.
  • Additional Ablation Studies (cSdP, s44D): cSdP pointed out that the original ablation studies did not sufficiently distinguish the contributions of different retrieval paths and IAR. s44D also requested further experimental evidence supporting the advantages of individual design choices. In response, we conducted additional ablations on retrieval strategies and ranking methods, validating the respective contributions of dual-path retrieval and IAR. The results have been incorporated into Section 4.3 and Appendix A.7.
  • Clarifications and Experiments on Retrieval Budgets (cSdP, s44D): cSdP raised concerns about context budget differences across memory representations. In response, we compared evidence units and atomic facts, finding comparable retrieval and QA performance. s44D raised questions about the scope of computational cost measurements, candidate pool sizes, and retrieval latency. To address these concerns, we clarified the scope of the cost measurements in Section 4.6 and further explained that the compared retrieval strategies use the same candidate pool configuration. We also conducted additional per-query latency analyses. The additional experiments and analyses have been incorporated into Appendix A.7.
  • Additional Failure Case Analysis (s44D): s44D noted that the limitations of the associative and temporal retrieval mechanisms were insufficiently discussed. In response, we added a failure case analysis illustrating how the two retrieval paths may interfere with each other. The corresponding analysis has been incorporated into Section 4.5 and Appendix A.9.

Overall, we have addressed the reviewers' main concerns through additional experiments and further analyses. We appreciate the follow-up feedback from s44D, who acknowledged that the additional experiments and failure case analysis addressed most of the concerns. Reviewer nhct also acknowledged the broader evaluation. We believe these efforts have strengthened the evaluation and presentation of CoRelMem. We sincerely thank all reviewers again for their time and valuable feedback.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.