acceptodds
Under review as a conference paper at ICLR 2027

OpCtx: Evolving Event Views for Budget-Constrained Agent Memory

Abstract

Long-term memory enables language agents to accumulate knowledge across interactions. However, as histories grow, answering new questions requires organizing observations of events that evolve across interactions and recovering useful evidence within limited reading budgets. These challenges are amplified in multimodal settings, where images contain overlapping details about objects, attributes, and relations. Existing memory pipelines still face two difficulties in addressing these challenges. First, they leave relationships among observations of the same event implicit, requiring the answering model to reconstruct these relationships from scattered evidence. Second, they retrieve multiple facts conveying the same information when selection relies primarily on similarity, leaving less of the retrieval budget for complementary evidence. Retrieving more candidates and applying model-based reranking can improve selection, but reranking costs grow with candidate text. To address these difficulties, we propose OpCtx, a memory architecture connecting online event organization with question-conditioned evidence reading. OpCtx organizes source-linked observations into cross-interaction event views and, using a training-free Reader, aggregates selected facts and supported event information into structured evidence. This helps the answering model interpret relationships and changes across interactions. To improve evidence coverage within the reading budget, a coverage-preserving selector balances lexical and semantic signals under explicit fact-count limits and an optional character budget. Experiments on MemEye, LoCoMo, and LongMemEval demonstrate the effectiveness of OpCtx across multimodal and text-only memory tasks.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Agent Memory, Multimodal Agent Memory, Budget-Constrained Retrieval

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_mJaWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Existing memory pipelines still face two difficulties in addressing these challenges. First, they leave relationships among observations of the same event implicit, requiring the answering model to reconstruct these relationships from scattered evidence. Second, they retrieve multiple facts conveying the same information when selection relies primarily on similarity, leaving less of the retrieval budget for complementary evidence. The authors propose OpCtx, a memory architecture connecting online event organization with question-conditioned evidence reading to address the challenges .

Critical strengths

S1: The method has strong empirical results across 3 benchmarks. Consistent gains over baselines on multi-hop (73.76% vs Mem0's 69.50% on LoCoMo) and temporal (83.18% vs 78.82%) questions. S2: Clean ablation design. The ablation (Full/w-o Reader/w-o Selector/w-o Online) with shared frozen Writer facts enables clean attribution. The component isolation is convincing. S3: The paper is clearly written and easy to follow. The figures are well-presented, and the overall pipeline is clearly designed.

Critical weaknesses

W1: One of the issues raised in the paper is the limited retrieval budget and suboptimal reranking. However, the experiments do not compare different reranking methods. Instead, they compare many end-to-end approaches. There is no comparison of rerankers such as MMR, cross-encoder, and greedy under the same candidate pool, reader, and budget, nor are there any budget–performance curves. Therefore, it remains unclear whether min-max selection truly outperforms existing methods. A direct comparison under controlled conditions would help strengthen this claim. W2: The paper's proposed approach to "scattered memories and cross-session evidence organization" has already been extensively explored in existing solutions, such as episodic memory, temporal knowledge graphs (KGs), and multi-hop RAG. The lack of mechanism-level comparison makes the proposed method appear as a recombination of existing modules, which weakens the perceived novelty of the core contribution. W3:The claimed advantages of the proposed "online event view" are not substantiated. The paper lacks analysis of temporal evolution, update latency, and noise in event extraction. Furthermore, the fact that removing the Online component leads to better performance in certain ablation settings raises questions about whether the dynamic mechanism is truly necessary.

Questions for authors

Q1: How does the event view count scale with session count on MemEye? Q2: Why does w/o Online improve MemEye Y3? Q3: See in weakness.

4 comments
AuthorsOpenReview ↗

Response to Reviewer mJaW: Controlled Comparisons and Event Organization

We sincerely thank the reviewer for the thoughtful comments and for recognizing our empirical results and controlled ablations. Your suggestions help us clarify the mechanism-level contribution and make the relevant experimental evidence easier to locate.

W1. Controlled selection comparisons and budget–performance curves

The main tables compare complete memory architectures. Appendix C.5, Table 11 compares selectors on MemEye with the same candidate pool, 30-fact limit, character ceiling, Reader, and QA settings, without model reranking:

Selection methodMemEye score (%)
Dense Top3043.66
BM25 Top3045.28
MMR Top3040.97
Equal-weight multichannel fusion45.55
OpCtx Selector47.44

This matched comparison gives OpCtx a 6.47-point advantage over MMR. Figure 3 and Appendix C.4 report budget–performance curves over final fact limits of 5–25 on LoCoMo and 10–50 on MemEye, without a reranker on MemEye. Both report QA performance and input tokens with memory, Reader rules, models, and prompts fixed.

Removing Qwen reranking on LoCoMo while keeping 20 final facts gives 80.06% accuracy with Selector and 77.40% without it, supporting Selector's independent contribution. We will highlight these controls and budget curves in the main text.

W2. Relationship to episodic memory, temporal KGs, and multi-hop RAG

We agree that structured, cross-session memory has substantial prior work. Our comparisons include HyperMem, HippoRAG 2, and Mem0. OpCtx couples evolving event views and evidence-preserving storage with budgeted selection and question-conditioned reading.

Event organization. HyperMem organizes topics, episodes, and facts in a hypergraph and updates topic summaries. OpCtx maintains typed event views over immutable source facts: members must describe the same occurrence, a continuing process, or an explicit state-update chain; a shared topic or entity is insufficient. Each view links claims to supporting fact IDs. Membership and claims evolve while original observations, including earlier and later visual states, remain unchanged.

Selection and reading. Visual observations may share an object or topic yet differ in relevant attributes or temporal states. OpCtx retains object anchors in attribute-specific queries, prioritizes query-term coverage, and balances normalized retrieval utilities within the evidence budget. Reader organizes selected facts, session summaries, and supported event claims for counting, temporal ordering, and comparison, including a claim only when all its supporting facts are selected.

Table 11 directly tests Reader: with identical facts, summaries, and event claims, OpCtx Reader scores 47.44%, versus 41.64% for a generic reasoning Reader. MemEye uses no model reranking. We will clarify the differences in grouping criteria, update units, selection objectives, and evidence construction.

W3 / Q2. Online event views and the Y3 result

The Y3 aggregate scores 44.79% with Online and 52.08% without it. Table 8 localizes this decline to X4Y3: Full scores 28.85%, versus 50.00% without Online. In the other three Y3 subcategories, Full matches or exceeds the corresponding scores without Online. Across all MemEye questions, Online improves the score from 37.47% to 47.44%.

Your observation about X4Y3 also helped us identify a limitation. We appreciate you pointing this out. These questions concern fine visual changes whose descriptions can be lengthy. Our inspection indicates that additional context introduced by Online competes with these descriptions within the evidence budget, displacing useful details. X4Y3 contains only 26 questions, so changes in a few answers noticeably affect the category score. We will clarify this limitation and prioritize detail preservation during selection and reading in future work.

Q1. How event-view count grows with session count

Section 3.2 and Appendix E.1 distinguish view count from evidence accumulated within a view. Create adds a view connecting current and historical evidence. Extend adds current evidence while retaining existing members, without increasing the view count. Repeated observations of the same event accumulate within a view; newly identified cross-session event relationships can create new views.

Updates retrieve historical facts and active views from the observed prefix and revise only the affected views and embeddings. Each update allows at most 8 operations, with at most 24 member facts per operation, so it can introduce at most eight new views. View growth depends on the mix of newly identified and continuing events, rather than a fixed number of views per session. Before committing an update, OpCtx checks source-reference visibility, claim-support membership, and operation validity. We will clarify these growth and update rules in the paper.

AuthorsOpenReview ↗

Follow-up on Our Rebuttal

Dear Reviewer,

Thank you again for your time and thoughtful review. We have responded to all your questions and concerns and provided additional experimental results. As the rebuttal period draws to a close, we would greatly appreciate your feedback and would be happy to clarify any remaining concerns. If our responses and additional results help resolve your concerns, we would be very grateful if you could reconsider your overall assessment.

We also noticed that “Flag for Ethics Review” was marked “Yes,” although no specific ethical concerns were mentioned in the review. Could you kindly double-check whether this selection was intended? If so, we would greatly appreciate further clarification so that we can address any concerns appropriately.

Dear Reviewer,

Thank you again for your time and consideration. With less than 30 hours remaining in the rebuttal period, we would be very grateful if you could review our responses and additional experimental results and share any further feedback. We would be happy to address any remaining concerns before the deadline.

Thanks for your response and additional experiments and I will keep my score.

Reviewer_aLpkOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

OpCtx organizes observations into evolving event views and combines budget-constrained evidence selection with a training-free Reader. Experiments on MemEye, LoCoMo, and LongMemEval-S show improved QA performance, with ablations assessing the contributions of its components.

Critical strengths

OpCtx achieves the best overall scores on all three benchmarks. It also performs well with two different LLMs in the text-only experiments.

OpCtx keeps the original facts and links related observations across sessions. This makes evidence easier to trace and changes over time easier to follow.

The Selector chooses evidence within the given budget and runs quickly on the tested MemEye candidate pools, with a median execution time of 0.187 seconds.

Critical weaknesses

The cost comparison leaves out the Reader and other earlier steps, so it is unclear whether OpCtx is cheaper or faster overall.

Randomizing links retrieves almost as many relevant sources as the original links (Appendix C.5, 86.84% vs. 88.83%). The paper does not report QA results with random links, so it is unclear how much correct event links help.

The paper does not explain how to fix facts assigned to the wrong event.

The Selector scores only 1.89 points higher than equal-weight fusion (Table 11). The paper does not show whether this gain is reliable or whether all parts of the more complex method are needed.

MMA and M2A are not clearly defined.

Questions for authors
  1. How many tokens and how much time does OpCtx use in total, including memory construction and the Reader? Does it still outperform other methods under the same total budget?

  2. How does QA accuracy change when event links are randomized, but other settings stay the same? Do correct event links improve the answers?

  3. Can OpCtx fix wrong event links? Why does adding Online lower the score on questions about fine visual changes from 50.00% to 28.85% (Table 8)?

5 comments
AuthorsOpenReview ↗

Response to Reviewer aLpk: End-to-End Cost, Event Links, and Selector Analysis

We sincerely thank the reviewer for the careful review and for recognizing our empirical results, evidence traceability, and Selector efficiency. Your comments have helped us clarify the method and strengthen the experimental analysis.

W1 / Q1. End-to-end cost, budget comparison, and latency

Embedding, reranking, and QA models differ in price, so we compare all stages in USD. Writing includes the Writer and Online updates; QA includes query planning, the Reader, and answer generation. Costs cover the complete 1,540-question LoCoMo and 500-question LongMemEval-S evaluations, including construction, reranking, and embeddings. Per-question costs are in US cents.

LoCoMo:

MethodWritingQARerankingEmbeddingsTotalPer Q (cents)
HyperMem6.971.1000.038.100.53
Mem08.881.7700.0210.670.69
MemoryOS10.751.2800.0412.070.78
RAG01.1100.041.150.07
RAPTOR0.291.0900.051.430.09
HippoRAG21.551.131.490.064.230.27
SeCom0.361.0400.031.430.09
MemGAS0.233.2000.143.570.23
OpCtx4.271.630.340.056.280.41

LongMemEval-S:

MethodWritingQARerankingEmbeddingsTotalPer Q (cents)
Full Context016.550016.553.31
RAG00.4207.918.331.67
Mem0228.370.2207.14235.7447.15
SimpleMem206.210.3006.79213.3042.66
OpCtx186.150.660.613.82191.2438.25

On LoCoMo, OpCtx achieves 81.75% accuracy at $6.28, versus HyperMem's 76.23% at $8.10 and Mem0's 76.10% at $10.67. On LongMemEval-S, OpCtx achieves 81.00% at $191.24, versus SimpleMem's 72.00% at $213.30 and Mem0's 50.40% at $235.74. OpCtx thus achieves higher accuracy at lower total cost.

LongMemEval-S has one question per history, making Full Context appear especially inexpensive. Repeated questions let OpCtx reuse memory and share construction costs, while Full Context supplies the full history each time.

LoCoMo answering latency:

MethodQA latency (s/question)
OpCtx2.842
HyperMem5.775
RAPTOR3.401
HippoRAG22.911
SeCom2.785
Mem05.374
SimpleMem3.291

These timings cover answering only; network delays and retries prevent consistent measurement of construction time.

W2 / Q2. Randomized event links: source recall and QA accuracy

In the random-link setting, connections and candidates were uncapped, linking more views and retrieving more facts. This explains its high candidate-source recall of 86.84%, versus 88.83% for Full. Many facts were discarded before reranking. On LoCoMo multi-hop, QA accuracy was 52.13% with randomized links and 73.76% for Full OpCtx with correct event links, a 21.63-point difference. High source recall thus did not translate into comparable QA accuracy. We will add these results to Appendix C.5.

W3 / Q3. Incorrect event assignments and fine visual changes

OpCtx maintains event membership separately from immutable source facts, so an incorrect association does not overwrite the original evidence. This separation allows corrections to target the event layer while preserving provenance. We will clarify this distinction in the paper.

X4Y3: Your observation about X4Y3 also helped us identify a limitation. We appreciate you pointing this out. Full OpCtx scores 28.85%, compared with 50.00% without Online. Online adds context that competes with lengthy descriptions of fine texture changes within the evidence budget, displacing useful details. This subset contains only 26 questions, so changes in a few answers can noticeably affect the score. We will clarify this limitation and prioritize detail preservation during selection and reading in future work.

W4. Selector effectiveness and component necessity

Beyond the 1.89-point improvement over equal-weight fusion in Table 11, supplementary ablations at a 50-fact budget show decreases of 0.40, 1.35, and 1.89 percentage points when removing query projection, removing coverage priority, and replacing max–min with summation, respectively. These results provide additional evidence for the individual design choices. We will also distinguish the observed improvements from claims of statistically reliable superiority.

W5. Definitions of MMA and M2A

MMA stands for Multimodal Memory Agent (Lu et al., 2026) and assesses evidence reliability through source credibility, recency, and memory agreement. M2A, Multimodal Memory Agent with Dual-Layer Hybrid Memory for Long-Term Personalized Interactions (Feng et al., 2026), combines an immutable raw-message store with a semantic-memory store. We will define and cite both at first mention.

Thank you for the additional results. I would appreciate clarification on Naive RAG’s chunking, retrieval method, top‑k, and reranking settings, as well as a comparison at the same end-to-end cost against a stronger RAG baseline that combines BM25 and dense retrieval and uses the same reranker as OpCtx.

For the randomized-link experiment, please report QA accuracy when only the links are randomized and the link and candidate limits are held fixed. Please also explain how OpCtx detects and corrects erroneous links after they are added. These unresolved issues still affect my overall score.

AuthorsOpenReview ↗

Follow-up Response to Reviewer aLpk: RAG Settings and Event Links

Thank you for your constructive follow-up. We appreciate the opportunity to clarify the retrieval settings and strengthen the analysis of event links.

1. Naive RAG settings and comparison with a stronger hybrid RAG baseline

The original Naive RAG uses whole sessions as chunks and dense retrieval, with top-5 sessions and a 3,500-token limit to keep QA input usage comparable to the other baselines. It does not use reranking. For the additional comparison, we increased the retrieval budget and combined dense retrieval and BM25, followed by the same reranker used by OpCtx. The costs below cover the full pipeline, including reranking.

On LoCoMo, the stronger hybrid RAG achieves 75.13% accuracy at $11.57, compared with 81.75% at $6.28 for OpCtx. Thus, OpCtx remains more accurate even when hybrid RAG is allowed a higher total expenditure.

On LongMemEval-S, hybrid RAG achieves 71.40% accuracy at $41.28, compared with 81.00% for OpCtx. The benchmark evaluates only one question per conversation history, assigning the entire memory-construction cost to a single answer and making memory systems appear more costly than in practical multi-query use. Under this protocol, even using the full context with reranking would still cost less than OpCtx. In real-world applications, the same memory typically serves multiple questions. The cost advantage for memory systems over full-context answering comes from building memory once and then retrieving a small amount of relevant information for subsequent questions, rather than repeatedly supplying the entire history. This spreads the construction cost across queries while reducing the input required for each answer.

2. Controlled randomized-link experiment

We randomized only the event links, keeping all other settings unchanged, including the link and candidate-count limits. Both configurations retrieved an average of 4.71 views per question, with 22.91 facts per view. QA input averaged 2,629 tokens per question with correct event links and 2,659 with randomized links.

On LoCoMo multi-hop, randomized-link QA accuracy was 59.22%, compared with 73.76% for Full OpCtx, a 14.54-point difference. The accuracy gap persists with matched average retrieval sizes and similar QA input token counts, supporting the contribution of correct event associations.

3. Erroneous links: validation, mitigation, and correction

OpCtx validates source-reference visibility, claim–support membership, and operation validity before accepting event updates. At query time, evidence undergoes question-conditioned selection, and the Reader includes an event claim only when all its cited supporting facts are present in the selected evidence. Dense and BM25 retrieval also provide independent access to the unchanged source facts.

These safeguards help limit the downstream influence of erroneous associations. The evaluated implementation does not automatically retract accepted links. At the evidence level, preserving source observations retains earlier states and preferences alongside later updates, helping reconstruct how an event or preference evolves across sessions without overwriting its history. We will clarify the distinction between error mitigation and link correction in the revision.

We would be grateful if you could reconsider your overall assessment in light of these additional results and clarifications.

Reviewer_BPu1OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies long term memory for language agents in multimodal and text only settings. It argues that existing pipelines leave relations among observations of the same event implicit, and that similarity based retrieval wastes the reading budget on redundant facts. The paper proposes OpCtx with three parts. A local Writer stores immutable source linked facts per batch, and an Online module creates or extends cross session event views over these facts. A Selector picks a fact set that first maximizes question term coverage and then maximizes the minimum normalized utility across dense and BM25 channels of object anchored queries, solved as a MILP. A training free Reader aggregates the selected facts, session summaries and supported event claims into an evidence packet for QA. Validation covers MemEye with 371 questions, LoCoMo with 1,540 non adversarial questions and LongMemEval-S with 500 questions. With Gemini 2.5 Flash, OpCtx reaches 81.75 on LoCoMo against 76.36 for Full Context, and 81.00 on LongMemEval-S against 72.00. On MemEye it reaches 47.44 against 39.62 for FC with visual input. Ablations remove Online, Reader and Selector, and Appendix D gives a retention property of the max min objective.

Critical strengths
  1. The empirical gains are large and consistent. OpCtx is the best method overall on all three benchmarks in Tables 1 to 3, and the ranking holds when the backbone changes from Gemini 2.5 Flash to DeepSeek V4 Flash, with 79.22 on LoCoMo and 79.00 on LongMemEval-S. The largest gains appear in multi hop, temporal, knowledge update and multi session questions. This matters to my assessment because these are the categories where memory organization should help, so the results are relevant to the stated problem and not only to easy lookups.
  2. The component study is more careful than usual. Online, Reader and Selector are ablated on shared frozen Writer facts, and Appendix C.5 adds source recall in Table 10, matched selection controls in Table 11, and a generic reasoning Reader control. Removing the Reader drops MemEye from 47.44 to 38.54, and the generic Reader gives 41.64. This supports that question conditioned evidence organization is a real source of the gain.
Critical weaknesses
  1. The role of the event structure itself is not established. Table 10 shows that removing Online drops LoCoMo multi hop candidate source recall from 88.83 to 67.30. However, randomizing the event links over 20 seeds still gives 86.84, only 1.99 points below Online. So most of the recall gain survives when the event links are random. This suggests that the benefit mainly comes from expanding the candidate pool through view routing, and not from correct event membership. The paper claims event conditioned organization as the first contribution, so this weakens the central claim. QA accuracy with random links is not reported. On MemEye, removing Online also gives a higher score on X4 Y3, 50.00 against 28.85 in Table 8.
  2. The budget and efficiency claims are not supported by the reported cost. The title and Section 3.3 frame OpCtx as budget constrained and as reducing reliance on model based reranking. In Appendix E.2, LoCoMo and LongMemEval-S still apply Qwen3-Reranker-4B after the Selector. On MemEye, the wide Qwen pipeline scores 49.19, higher than the 30 fact Selector at 47.44 in Figure 4c, and the 30 fact Selector is only 1.89 points above equal weight fusion at 45.55 in Table 11. Table 13 shows 3,174.92 QA input tokens per question for OpCtx on LoCoMo, more than SimpleMem at 1,568 and LightMem at 982. The cost of the Writer, the Online update model, query extraction and the Reader is excluded, as Section 5 and Appendix B.3 admit. A method with several extra LLM stages per batch and per question cannot be judged as budget efficient without end to end cost.
Questions for authors
  1. Please report QA accuracy on LoCoMo and MemEye with randomized event links, using the same setup as Table 10. If QA accuracy stays close to Online, the event structure is not the cause of the gain. This answer could change my rating.
  2. Please report end to end cost per question and per memory batch for OpCtx and the main baselines, including LLM calls and tokens for the Writer, Online updates, query projection, Reader and QA, together with wall clock time.
  3. On LoCoMo and LongMemEval-S, the full pipeline uses the Qwen reranker after the Selector. Please clarify what the main results look like without the reranker, so the contribution of the Selector can be judged on all three benchmarks.
  4. On MemEye, the text memory baselines use the image descriptions given by the benchmark, while OpCtx processes images with Gemini 2.5 Flash-Lite, as stated in Appendix B.2. Please clarify whether OpCtx with the same benchmark descriptions keeps its lead over FC and SRAG with text input.
  5. The MemEye mapping in Appendix B.3 gives 0.5 credit to a raw judge score of 0.25. Please clarify whether this mapping follows the MemEye protocol, and report scores with the raw judge values.
  6. Each answer receives one judge verdict and no variance is given. Please report results over several runs or seeds, at least for the ablations in Figure 4 where several gaps are below 3 points.
  7. Please clarify how the baselines were tuned. Several reproduced numbers are low, for example SeCom at 13.08 on LoCoMo temporal and Naive RAG at 16.40 on LongMemEval-S with Gemini 2.5 Flash. I suggest comparing them with the numbers in the original papers where the settings overlap.
3 comments
AuthorsOpenReview ↗

Response to Reviewer BPu1: Additional Experiments and Evaluation Clarifications

We sincerely thank the reviewer for the careful evaluation and constructive suggestions. Your feedback has helped us strengthen the experimental analysis and clarify several important aspects of our work.

1. Random links and X4Y3

The random-link setting did not cap connections or candidate counts, connecting more views and retrieving more facts. This explains the high candidate-source recall. Many facts were then discarded before reranking; LoCoMo multi-hop QA accuracy was 52.13%, versus 73.76% for Full. We will add these QA results to Appendix C.5.

Your observation about X4Y3 also helped us identify a limitation. We appreciate you pointing this out. Online adds context that competes with lengthy descriptions of fine texture changes within the evidence budget, displacing useful details. With only 26 questions, a few changed answers noticeably affect the score. We will clarify this limitation and prioritize detail preservation during selection and reading in future work.

2. End-to-end cost

Writing includes Writer and Online updates; QA includes query planning, Reader, and answer generation. Costs cover all 1,540 LoCoMo and 500 LongMemEval-S questions, including construction. Stage costs and totals are USD; ¢/Q is US cents per question.

LoCoMo:

MethodWriteQARerankEmbedTotal¢/Q
HyperMem6.971.1000.038.100.53
Mem08.881.7700.0210.670.69
MemoryOS10.751.2800.0412.070.78
RAG01.1100.041.150.07
RAPTOR0.291.0900.051.430.09
HippoRAG21.551.131.490.064.230.27
SeCom0.361.0400.031.430.09
MemGAS0.233.2000.143.570.23
OpCtx4.271.630.340.056.280.41

LongMemEval-S:

MethodWriteQARerankEmbedTotal¢/Q
Full Context016.550016.553.31
RAG00.4207.918.331.67
Mem0228.370.2207.14235.7447.15
SimpleMem206.210.3006.79213.3042.66
OpCtx186.150.660.613.82191.2438.25

OpCtx costs less than HyperMem, Mem0, and MemoryOS on LoCoMo, and than Mem0 and SimpleMem on LongMemEval-S. Because each history has only one question, this evaluation makes Full Context appear especially inexpensive and does not capture the benefit of memory reuse. Repeated queries can share OpCtx's construction cost, while Full Context supplies the full history each time.

3. Without reranking

Scores below are percentages.

LoCoMo retains the same final limit of 20 facts:

SettingSingle-hopMulti-hopTemporalOpen-domainOverall
Full86.6873.7683.1857.2981.75
w/o Qwen85.6166.3185.6753.1280.06
w/o Qwen & Selector83.7160.9982.8752.0877.40

With Qwen removed, further ablating Selector lowers LoCoMo accuracy from 80.06% to 77.40%, showing a 2.66-point gain from Selector.

LongMemEval-S:

SettingKUMSSSASSPSSUTROverall
Full91.0375.1975.0073.3394.2978.2081.00
w/o Qwen85.9073.6876.7963.3390.0072.9377.40

Without Qwen, overall accuracy drops by 1.69/3.60 points on LoCoMo/LongMemEval-S but still exceeds all reported Gemini 2.5 Flash baselines. LongMemEval-S achieves 387/500 correct, 5.40 points above its strongest baseline.

4. Same benchmark captions

OpCtx accepts images directly. We use Gemini 2.5 Flash-Lite to read images because the benchmark-provided image descriptions are not well suited to building event views in OpCtx. Using the same captions as FC(T)/SRAG(T), without images, OpCtx scores 34.23 mapped/32.21 raw, retaining its lead over FC(T) by 4.04/3.10 points. The margin is smaller than with direct image input.

5. MemEye scoring

The mapping follows MemEye: raw {0,0.25,0.5,0.75,1} maps to {0,0.5,0.5,1,1}.

Key comparisons:

MethodMappedRaw
FC(T)30.1929.11
SRAG(T)23.4522.57
FC(V)39.6238.07
SRAG(V)35.8534.23
MMA37.8736.39
OpCtx47.4444.95

OpCtx remains first and all rankings are unchanged. We will report both scores and the mapping explicitly.

6. Three seeds

We reran the whole experiment with three seeds. Values are mean ± sample SD, in percentage points.

LoCoMo: Full 81.73±0.16; w/o Reader 78.96±0.24; w/o Selector 78.44±0.11; w/o Online 76.43±0.17. MemEye: Full 47.35±0.28; w/o Reader 38.50±0.08; w/o Online 37.29±0.21.

Full exceeds each tested ablation in all three runs, supporting repeatable component gains.

7. Baselines

SeCom uses its official GitHub implementation and default parameters without further tuning. Repeated evaluations consistently produced similarly low temporal scores. Some studies have also reported similarly low temporal scores for SeCom.

Naive RAG retrieves whole sessions on LongMemEval-S. Long sessions limit how many fit under the same QA token budget as other baselines, reducing coverage of evidence spread across sessions and helping explain its low score.

AuthorsOpenReview ↗

Follow-up on Our Rebuttal

Dear Reviewer,

Thank you again for your time and thoughtful review. We have responded to all your questions and concerns and provided additional experimental results. As the rebuttal period draws to a close, we would greatly appreciate your feedback and would be happy to clarify any remaining concerns. If our responses and additional results help resolve your concerns, we would be very grateful if you could reconsider your overall assessment.

Dear Reviewer,

Thank you again for your time and consideration. With less than 30 hours remaining in the rebuttal period, we would be very grateful if you could review our responses and additional experimental results and share any further feedback. We would be happy to address any remaining concerns before the deadline.

Comments on the paper

1 comment
Area_Chair_V4NROpenReview ↗

Please Engage in the Author-Reviewer Discussion

Dear Reviewers,

The authors have posted their responses to your comments and questions. Please review their replies and share any follow-up comments or questions, especially if any concerns remain unresolved or need further clarification.

Your engagement is important to ensuring a constructive and productive discussion. Thank you for your time and contributions to the ICLR review process.

Best regards,
The AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.