acceptodds
Under review as a conference paper at ICLR 2027

ContextFact: Remembering Facts in Context for Long-Term Conversational Agents

Abstract

Agents built on large language models must maintain continuity across repeated interactions with users. However, long-term conversations scatter facts across turns and sessions, making it difficult to retrieve the right evidence with enough context. Memory for such conversations must therefore preserve facts with their context while keeping evidence selectively accessible. We first study how memory construction and retrieval-unit boundaries affect answerability and evidence access. Reading the entire memory bank reveals gaps that cannot be resolved by retrieval selection alone, showing that some context must be preserved during memory construction. Adjacent-turn grouping improves evidence access at matched unique-turn budgets, while wider windows do not consistently help. This motivates keeping fact-specific context local while organizing cross-session evidence separately. We propose ContextFact, which links context-preserving fact memories with source-grounded entity representations. Within sessions, it gathers the context needed to interpret each fact. Across sessions, it aligns recurring referents and revisits the source dialogue to reconstruct coherent entity representations linked to the supporting facts. At query time, ContextFact retrieves fact and entity views, reranks them independently, and uses entity-linked expansion to surface related facts; complete entity representations then provide organized cross-session evidence. With Qwen3.5-27B as the backbone, ContextFact achieves state-of-the-art performance while using only of the evidence tokens consumed by the strongest evaluated baseline, obtaining overall judge accuracy on LoCoMo and on LongMemEval-S. Controlled ablations support context-preserving fact boundaries and source-grounded entity reconstruction, while entity readout improves multi-hop accuracy beyond adding more facts.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM Agents, Long-Term Conversational Memory, Context Preservation, Entity-Centered Retrieval, Multi-Hop Reasoning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_wM55OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper is to tackle the problem of how to construct long-term memory in a conversation to LLM agent. There are two major challenges when building the long-term memory in the existing systems: (1) the loss of important details and contextual information during memory construction, and (2) the difficulty of efficiently retrieving relevant evidence scattered across multiple dialogue turns and sessions. The authors aim to build a memory system that preserves the context necessary to understand individual facts while enabling efficient retrieval of related evidence across sessions.

To achieve this goal, the authors propose ContextFact framework, a two-view memory framework. The main contributions include: (1) conducting diagnostic experiments to investigate how memory construction and retrieval granularity affect answerability and evidence accessibility; (2) proposing context-preserving fact memories that retain relevant information across nonadjacent dialogue turns; and (3) introducing source-grounded entity representations that connect related facts across sessions and support entity-linked retrieval and reasoning.

The authors evaluate ContextFact on two long-term conversational memory benchmarks, LoCoMo and LongMemEval-S, against several existing memory systems. They also conduct controlled experiments and ablation studies to validate the effectiveness of the proposed components. Overall, the experimental results show that ContextFact achieves better QA accuracy than the evaluated baselines while using significantly fewer evidence tokens.a

Critical strengths
  • The comprehensiveexperimental evaluation is one of the strengths of this paper. The authors compare ContextFact with many existing baselines and evaluate it using different LLM backbones. They also test on both LoCoMo and LongMemEval-S, and include detailed ablation studies to validate different components.

  • The proposed method achieves the best or second-best results on many tasks in Table 1, which demonstrates its effectiveness. What's more, the method uses significantly fewer evidence tokens than the strongest baseline.

  • I like the idea of preserving the context of individual facts and connecting related information across sessions through entity representations. The combination of these two memory views is a good design choice, and the experimental results show that it works well in practice. For example, the entity view helps improve multi-hop QA accuracy from 76.95% to 84.04% on LoCoMo with Qwen3.5-27B.

Critical weaknesses
  • The technical novelty of this paper is somewhat limited. The proposed framework mainly combines existing techniques, such as context-aware memory extraction, entity linking, graph-based memory construction, and hybrid retrieval with reranking. While the combination works well, I do not see a fundamental difference from existing structured or hierarchical memory systems, such as Memora and Zep.

  • Although the proposed method achieves the best overall results, the improvements over the strongest baseline are relatively small. For example, the improvement is 2.20 percentage points on LoCoMo and only 1.80 points on LongMemEval-S with Qwen3.5-27B. I would like to see statistical significance tests or confidence intervals to better understand how reliable these improvements are.

  • The entity view does not always improve performance. For example, on open-domain questions in LoCoMo, ContextFact-E performs worse than ContextFact-F with both GPT-4.1-mini (59.38% vs. 64.58%) and Qwen3.5-27B (51.04% vs. 54.17%). I find this somewhat surprising because the entity view is expected to provide additional evidence rather than hurt the results.

Questions for authors
  • How would ContextFact handle new conversation sessions in a continuously running agent? In particular, does the source-grounded entity reconstruction need to be repeated when new information arrives? I am curious about the efficiency and scalability of this process.

  • Why does adding the entity view sometimes lead to worse results, especially on open-domain questions? If the additional entity information does not conflict with the original ContextFacts, I would expect the performance to at least remain similar. Could this be caused by irrelevant information, redundancy, or changes in the retrieved facts? It would be helpful to see some failure-case analysis.

5 comments

We thank you for recognizing the experimental evaluation, evidence efficiency, and complementary roles of the two memory views. We address your concerns and questions below point by point, along with the corresponding revisions in the manuscript (highlighted in blue).


W1. Contribution beyond existing memory approaches

"The technical novelty of this paper is somewhat limited. ... I do not see a fundamental difference from existing structured or hierarchical memory systems, such as Memora and Zep."

We would like to clarify how ContextFact differs from Memora, Zep, and related memory systems. The distinction concerns the criterion for defining independently retrievable facts and the source-grounded reconstruction of cross-session evidence for the reader.

First, we formulate and validate an explicit principle for memory-unit boundaries: memory boundaries should follow interpretive dependencies while keeping distinct facts independently retrievable. ContextFact implements this principle by grouping each focal fact with the context needed to interpret it, including information from nonadjacent turns. This specifies which information should remain together to balance selective retrieval and contextual completeness.

Our diagnostic experiments motivate this principle (lines 179-184). Under matched evidence-token budgets, ContextFact-F achieves the highest overall judge accuracy among the evaluated memory systems at every tested budget on LoCoMo. Comparisons with sentence-level and atomic decompositions of the same memory bank further support the resulting unit design (lines 366-397; Tables 9-10).

Second, we reconstruct the entity view from the source dialogue after cross-session alignment. Memora connects concrete memories through primary abstractions and cue anchors, while Zep represents episodes, entities, relations, and temporal facts in a knowledge graph. ContextFact revisits the dialogue associated with aligned referents to reassess provisional groupings and reconstruct coherent accounts of entities' aspects and developments. These accounts retain links to supporting facts and are supplied directly to the reader (lines 772-782).

The entity view therefore participates in both evidence discovery through linked facts and evidence organization through reconstructed accounts. Controlled ablations show that fact-plus-entity readout outperforms additional-fact controls on multi-hop questions, even when those controls supply more evidence tokens (Table 3; lines 422-431).

W2. Reliability of the reported improvements

"Although the proposed method achieves the best overall results, the improvements over the strongest baseline are relatively small. ... I would like to see statistical significance tests or confidence intervals to better understand how reliable these improvements are."

To assess variability in the reported results, we conducted three runs for each method and dataset. All accuracy columns report the percentage of answers judged correct, summarized as mean ± sample standard deviation across three runs. The following results use Qwen for both memory construction and answer generation, with GPT-4o-mini as the judge.

1. LoCoMo: full-pipeline reruns.

We conducted three independent runs of the full pipeline for ContextFact and Memora (P), rebuilding the memory bank and repeating answer generation and LLM judging in each run. We used the same configuration as in our main experiments. These results capture variation across memory construction, answering, and judging.

MethodOverallSingle-hopTemporalMulti-hopOpen-domain
Memora(P)84.68 ± 0.6890.57 ± 0.5682.04 ± 1.2681.32 ± 1.4351.74 ± 1.59
ContextFact-F84.70 ± 0.8090.05 ± 0.5487.23 ± 1.3676.48 ± 1.0853.47 ± 0.60
ContextFact-E86.41 ± 0.1090.41 ± 0.2589.62 ± 0.3682.86 ± 1.1451.04 ± 1.04

2. LongMemEval-S.

We kept each method's constructed memory bank fixed and performed three runs of answer generation and LLM judging. These results capture variation in answering and judging conditional on the constructed memories.

MethodOverallSingle-session UserSingle-session AssistantMulti-sessionTemporal ReasoningKnowledge UpdateSingle-session Preference
Memora(P)89.20 ± 0.2098.10 ± 0.8298.21 ± 0.0080.95 ± 0.4385.96 ± 1.5794.44 ± 0.7488.89 ± 5.09
ContextFact-F87.53 ± 0.5096.19 ± 0.8296.43 ± 0.0080.45 ± 0.0084.71 ± 1.5791.03 ± 0.0085.56 ± 1.92
ContextFact-E91.20 ± 0.2099.05 ± 1.6598.81 ± 2.0683.46 ± 0.0089.72 ± 1.1594.87 ± 2.2290.00 ± 0.00

On LongMemEval-S, ContextFact-E's mean overall accuracy exceeds Memora (P)'s by 2.00 percentage points, with a run-to-run standard deviation of 0.20 points for each method.

Q1. Online memory updates and reconstruction costs

"How would ContextFact handle new conversation sessions in a continuously running agent? In particular, does the source-grounded entity reconstruction need to be repeated when new information arrives? I am curious about the efficiency and scalability of this process."

We agree that online memory updating is important for long-running agents, which must incorporate new information while maintaining coherent memories across sessions. ContextFact already provides a starting point through incremental fact insertion and cross-session entity binding. For each new session, it extracts ContextFacts and entity entries, appends the facts to the fact bank, and matches the entries to existing provisional entities or creates new ones. This process does not require repeating fact extraction over earlier sessions (lines 245-252; Algorithm 1, steps 3-8).

Regarding whether source-grounded reconstruction needs to be repeated, yes—in the current method, incorporating new sessions into the final entity view requires running reconstruction again using the updated source evidence. An online extension could reduce both how often reconstruction runs and how much source evidence it revisits.

To reduce reconstruction overhead while keeping new information accessible, we could separate immediate incorporation of new facts from deferred reconstruction of entity records. Newly extracted facts could become available for retrieval while their source turns accumulate in buffers associated with the affected entities. Reconstruction could then be triggered by accumulated evidence or a detected inconsistency, allowing multiple updates to be processed together.

A second consideration is to localize reconstruction while preserving access to source evidence. An update could reconsider the affected entity and, when identity decisions require it, its local review group. Reconstruction could use the current records and newly buffered turns, with provenance links providing access to earlier dialogue when needed. This would allow previous interpretations to be reconsidered as new information arrives. Batching updates and restricting their review scope could reduce repeated reconstruction work while keeping recent facts accessible.

Scalability also calls for compression and pruning to control the growth of active memory. An online extension could consolidate overlapping records, remove redundant entries, and archive older records while retaining their provenance for later retrieval. Each reconstruction call could select relevant records and source evidence within an explicit token budget. This would keep update inputs manageable while preserving access to historical evidence needed for temporal questions or identity correction.

Evaluating this extension would require studying reconstruction triggers, compression and retention policies, and consistency after entity merges or splits. These choices must balance long-term answer accuracy, memory growth, and update cost. We regard reliable online memory updating as an important future direction, building on the incremental binding and source provenance already present in ContextFact.

Reviewer_we3VOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes ContextFact, a dual view memory framework that preserves the interpretive context of each fact and links facts across sessions via source grounded entity representations. At query time, dual view retrieval with a single entity-linked expansion and independent reranking supplies both focused facts and organized entity evidence to the reader.

Critical strengths

1.Well-motivated and clearly scoped design: Building on known memory problems, the paper abstracts a clear design principle: memory boundaries should follow interpretive dependencies while keeping evidence independently retrievable, and the resulting design is logically coherent and well developed. 2.The method is targeted and effective: it is organized around entities, can collect context across non-adjacent turns, and remains independently retrievable in real time. It directly answers the question raised in this paper rather than mechanically stacking modules.

Critical weaknesses

1.A detailed analysis of construction costs at different stages could be provided, including Cross-session Entity Binding, Topic-Based Entity Grouping, Source-grounded Entity Reconstruction. Such a more in-depth analysis would allow for further cost estimation, thereby improving the estimation of contextfact's actual deployment costs.

2.Experiments with different models for memory construction could be provided to evaluate the impact of different models on construction, and cross experiments could also be added, such as constructing memory with Qwen3.5-27B and evaluating with GPT-4.1-mini, to verify cross-model transferability. In addition, the paradigm of weak model construction with strong model reasoning could be further explored to reduce the overall system cost.

Questions for authors

see above

3 comments

We thank you for recognizing the memory-boundary design principle and the coherence of the framework. We address your concerns and questions below point by point, along with the corresponding revisions in the manuscript (highlighted in blue).


W1. Construction costs by stage

"A detailed analysis of construction costs at different stages could be provided, including Cross-session Entity Binding, Topic-Based Entity Grouping, Source-grounded Entity Reconstruction."

We have added a stage-by-stage breakdown of construction token usage on LoCoMo for Qwen3.5-27B and GPT-4.1-mini. We report this breakdown in Table 17 and discuss the construction stages in lines 1983-1988. All counts are averaged per conversation and reported in thousands of tokens (k = 1,000). The percentages below are calculated from input and output tokens combined.

Construction stageQwen Input (k)Qwen Output (k)GPT-4.1-mini Input (k)GPT-4.1-mini Output (k)
Context-preserving fact extraction132.234.4122.730.8
Cross-session entity binding92.114.789.723.3
Topic-based entity grouping5.63.83.22.0
Source-grounded entity reconstruction66.521.757.915.0
Total296.574.6273.571.1

Fact extraction accounts for approximately 45% of total construction tokens for both models. Cross-session entity binding is the largest component of entity construction, accounting for 29-33% of total construction tokens, followed by source-grounded reconstruction at 21-24%. Topic-based grouping accounts for 1-3%.

W2. Construction-reader transfer and model choice

"Experiments with different models for memory construction could be provided ... to verify cross-model transferability. In addition, the paradigm of weak model construction with strong model reasoning could be further explored to reduce the overall system cost."

We have added a separate construction-reader evaluation on LoCoMo, testing all 25 pairings of five construction models and five readers for both ContextFact-F and ContextFact-E. Within each variant, we vary the construction and reader models while keeping the remaining experimental settings unchanged.

The table below reports ContextFact-E's overall accuracy: the percentage of answers judged correct by GPT-4o-mini. Rows identify the construction model, and columns identify the reader. Ranges are maximum minus minimum accuracy, measured in percentage points (pp). Bold entries mark pairings that use the same model for construction and reading (Table 19).

Construction ↓ / Reader →Qwen3.5-27BGPT-4.1-miniSol 6.1Sonnet 5.0Opus 5.5Range over readers (pp)
Qwen3.5-27B86.3086.4388.3889.8793.577.27
GPT-4.1-mini85.8486.7588.9690.0693.387.54
Sol 6.185.5887.4087.8688.1893.648.06
Sonnet 5.086.4987.0188.7789.9493.316.82
Opus 5.586.1787.2188.5189.9493.967.79
Range over construction models (pp)0.910.971.101.880.65

1. How much does the construction model affect accuracy?

For a fixed reader, accuracy varies by 0.65-1.88 percentage points across construction models (last row). For a fixed memory bank, it varies by 6.82-8.06 points across readers (last column). In these experiments, accuracy therefore varies more with reader choice than with construction-model choice.

2. Can a reader use memories constructed by another model?

GPT-4.1-mini achieves 86.43% with Qwen3.5-27B-built memories, compared with 86.75% with its own. The reader ordering also remains identical across all five construction models in the ContextFact-E comparison. These results support cross-model use of the constructed memories.

3. Can Qwen construct memories for a stronger reader?

Opus 5.5 achieves 93.57% with Qwen3.5-27B-built memories, compared with 93.96% when Opus also constructs the memories. Starting from Qwen for both construction and reading (86.30%), replacing only the constructor with Opus changes accuracy by −0.13 percentage points, whereas replacing only the reader adds 7.27 points. This supports using Qwen for memory construction while reserving Opus for answering.

For ContextFact-E, replacing the Qwen reader with Opus produces the largest category-level gains on multi-hop and open-domain questions across all five construction models (lines 2047-2051). ContextFact-E also improves overall accuracy over ContextFact-F in all 25 construction-reader pairs. Averaged equally across these pairs, the gains are 1.1 percentage points overall and 3.9 points on multi-hop questions (lines 2083-2087).

The complete results are reported in the revised manuscript, where Section 5.4 summarizes the construction-reader comparison, and Appendix F.9 details the protocol and accuracy variation.


Thank you again for your time and constructive feedback.

Thank you for your response.

Reviewer_hVT9OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces ContextFact, a framework for answering questions over long conversation histories. It constructs context-preserving fact memories and links them to entity representations reconstructed from the original dialogue. At query time, it retrieves both views, expands through their links, and independently reranks facts and entities before answering. Experiments on LoCoMo and LongMemEval-S report improved answer accuracy with substantially shorter reader evidence

Critical strengths
  1. Source-grounded entity reconstruction provides a useful correction step, revisiting the original dialogue to get a second chance to extract entity-grouping errors lost during extraction.
  2. Clear gain of entity-based memory on multi-hop questions
Critical weaknesses
  1. The paper seems incremental, which does not clearly establish a clear advance over closely related contextual and structured-memory approaches.
  2. Failure analysis of entity reconstruction is limited; it remains unclear how often the correction stage fixes existing errors or introduces new ones.
Questions for authors
  1. Is the entity view a general contribution? Adding it on top of another system's memory units, e.g., EMem's event units.
  2. How often does reconstruction merge, split or correct entities, and how often does it introduce errors such as wrong merges or records not supported by the source?
  3. Where does the entity view hurt? Open-domain accuracy drops when it is added: 64.58 to 59.38 with GPT-4.1-mini and 54.17 to 51.04 with Qwen (Table 1). How many questions flip from correct to wrong when entities are added, and are those flips caused by reconstruction errors?
5 comments

We thank you for recognizing the value of source-grounded entity reconstruction and the gains on multi-hop questions. We address your concerns and questions below point by point, along with the corresponding revisions in the manuscript (highlighted in blue).


W1. Contribution beyond existing memory approaches

"The paper seems incremental, which does not clearly establish a clear advance over closely related contextual and structured-memory approaches."

We would like to clarify how ContextFact differs from related contextual and structured-memory approaches, particularly in how it defines independently retrievable units and constructs cross-session evidence for the reader.

First, we formulate and validate an explicit principle for memory-unit boundaries: memory-unit boundaries should follow interpretive dependencies while keeping distinct facts independently retrievable. The key distinction is what determines a memory unit's boundary. SeCom uses contiguous topical segments, while EMem constructs event-centered units. ContextFact determines each unit's boundary from the context needed to interpret a focal fact, gathering supporting information that may be scattered across distant points in the conversation while keeping distinct facts separately retrievable (lines 764-770). This principle specifies which information should be preserved together to support both retrieval and interpretation. Our diagnostic experiments motivate it (Section 3), while budget-controlled comparisons and same-bank decomposition experiments support the effectiveness of the resulting units (Section 5.2 and Tables 9-10).

Second, we introduce source-grounded entity reconstruction to organize cross-session evidence for direct use by the reader. The distinction lies in reconstructing the entity view from the source dialogue after cross-session alignment. Zep represents episodes, entities, relations, and temporal facts in a knowledge graph, while Memora connects concrete memories through abstractions and cue anchors. ContextFact revisits the dialogue associated with aligned referents to reassess provisional groupings and reconstruct coherent accounts of entities' aspects and developments. These accounts retain links to supporting facts and are supplied directly to the reader (lines 772-782). Controlled ablations show that fact-plus-entity readout outperforms additional-fact controls, particularly on multi-hop questions, even when those controls supply more evidence tokens (lines 423-431 and Table 3).

Q1. Transfer across memory representations

"Is the entity view a general contribution? Adding it on top of another system's memory units, e.g., EMem's event units."

Yes. We have added experiments that apply our entity pipeline to EMem, SeCom, and SimpleMem, using both Qwen3.5-27B and GPT-4.1-mini. We retain each system's original memory bank, construct entity representations over its memory units, and apply our entity-linked retrieval and readout pipeline (+Entity). To account for differences in evidence length, we include +Budget controls. These retain each baseline's native retrieval pipeline and retrieve more original memory units to approximately match +Entity's mean evidence length.

The table below reports the Qwen3.5-27B results. All accuracy columns report the percentage of answers judged correct by GPT-4o-mini. Specifically, the question-type columns report accuracy within each category, and Overall reports accuracy across all questions. The Tokens column reports the mean number of memory-text tokens supplied to the reader per question.

MethodSettingTokensSingle-hopTemporalMulti-hopOpen-domainOverall
EMemBaseline1,44982.4077.2668.4450.0076.75
+Budget3,29485.8576.3278.3751.0480.32
+Entity view3,19386.5681.3181.2151.0482.27
SeComBaseline2,05577.7655.4560.9943.7567.92
+Budget4,84485.7365.7371.6354.1777.01
+Entity view4,87291.8081.3181.9148.9685.13
SimpleMemBaseline1,52976.6983.4971.6345.8375.26
+Budget2,96877.7685.0575.5347.9277.01
+Entity view2,86378.6087.8578.3744.7978.38

Across all three systems and both models, +Entity improves overall accuracy over both the original and +Budget baselines. With Qwen3.5-27B, its gains over +Budget range from 2.84 - 10.28 percentage points on multi-hop questions and 2.80 - 15.58 points on temporal questions. These results support transferring our entity pipeline to memory banks constructed by other systems at comparable mean evidence lengths. Results for both models are provided in Table 18 and summarized in Table 4.

W2 & Q2. Reconstruction outcomes and errors

W2. "... it remains unclear how often the correction stage fixes existing errors or introduces new ones."

Q2. "How often does reconstruction merge, split or correct entities, and how often does it introduce errors such as wrong merges or records not supported by the source?"

We have quantified reconstruction's grouping changes, audited reconstruction errors among downstream regressions, and compared reconstructed records with their underlying evidence. Among the 50 audited Qwen regressions (i.e., answers changing from correct to incorrect when the entity view is added), three were clearly attributable to reconstruction errors. We describe the grouping statistics and the two analyses below.

1. How often does reconstruction rewrite, merge, or split entities?

On the Qwen-built LoCoMo bank, extraction produces 1,901 entity entries, which cross-session binding consolidates into 1,412 provisional entities. Reconstruction produces 1,148 entities:

  • 961 provisional entities (68.1%) are rewritten one-to-one.
  • 441 provisional entities (31.2%) are merged into 165 entities.
  • 10 provisional entities (0.7%) are split into 22 entities.

The percentages use the 1,412 provisional entities as the denominator. All resulting records are reconstructed from their linked source dialogue, including those whose grouping remains one-to-one. Reconstruction therefore reassesses both entity groupings and the content of their records.

Quantifying how often reconstruction corrects or introduces errors would require a separate manual comparison of records before and after reconstruction against their source dialogue. Defining an error rate also requires distinguishing acceptable compression from omissions or distortions that change the meaning. QA accuracy alone cannot isolate reconstruction quality, since it also depends on retrieval and reader behavior. Thus, we conducted the two analyses below to examine how reconstruction affects downstream answers: an audit of reconstruction-related QA regressions and a comparison of reconstructed records with their underlying facts and source turns.

2. What reconstruction errors appear in the observed QA regressions?

We define reconstruction errors as statements in entity records that conflict with the source dialogue. We audited all 50 questions, spanning all four question types, that Qwen answers correctly with ContextFact-F but incorrectly with ContextFact-E. For each case, we checked the entity records supplied to the reader for conflicts with their source dialogue. Three regressions were clearly attributable to reconstruction errors, one each among single-hop, multi-hop, and temporal questions:

  • A record attaches "immersive" to the wrong aspect of a game.
  • A record dates the opening of an online store to an earlier advertising campaign.
  • A record describes a screening as the third when the dialogue mentions only two.

The remaining 47 regressions could not be clearly attributed to reconstruction errors. None of the 3 identified cases is an open-domain question. Appendix H.2 summarizes the patterns in the nine open-domain regressions (lines 2477-2516), and Appendix H.3 presents the failure case studies.

3. How do reconstructed records compare with their underlying evidence?

We also examine the downstream effect of the reconstruction. Holding the selected entities fixed, we replace their records with either their linked ContextFacts or their source dialogue turns. All accuracy columns report the percentage of answers judged correct. The Tokens column reports the average number of memory-text tokens supplied to the reader per question.

ModelReadoutTokensSingle-hopTemporalMulti-hopOpen-domainOverall
Qwen3.5-27BReconstructed records2,90790.4988.7984.0451.0486.49
Linked ContextFacts4,23591.0889.4182.9851.0486.75
Source turns7,88391.5688.7982.9847.9286.69
GPT-4.1-miniReconstructed records4,13991.4489.7281.5659.3887.27
Linked ContextFacts7,43690.4988.4780.8560.4286.43
Source turns13,24292.7586.6080.1457.2986.95

With Qwen3.5-27B, overall accuracy ranges from 86.49% to 86.75% across the three readouts. Linked ContextFacts and source turns improve overall accuracy by 0.26 and 0.20 percentage points, respectively, while source turns require 2.7 times as many tokens as reconstructed records. With GPT-4.1-mini, reconstructed records achieve the highest overall accuracy, 87.27%, using less than 1/3 of the source-turn tokens. Reconstructed records also achieve the highest multi-hop accuracy with both readers. These results support their usefulness as compact, organized evidence for answering questions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response

We thank all three reviewers for their thoughtful evaluations and constructive feedback. We appreciate hVT9's recognition of source-grounded reconstruction as a useful correction step and the multi-hop gains. We also appreciate we3V's positive assessment of the memory-boundary principle and the well-motivated, coherent design. We thank wM55 for recognizing the two-view design, comprehensive evaluation, and evidence efficiency.


We have provided point-by-point responses and revised the manuscript, with manuscript changes highlighted in blue. The main additions and clarifications in the manuscript and our responses are:

  • Clarifying our contributions (hVT9, wM55). We clarify two contributions: defining independently retrievable facts by their interpretive dependencies, and source-grounded entity reconstruction after cross-session alignment. Controlled granularity and readout comparisons support these choices (Appendix A; Sections 5.2-5.3).

  • Transfer across memory representations and models (hVT9, we3V). The entity pipeline improves overall accuracy across EMem, SeCom, and SimpleMem at comparable mean evidence lengths (Tables 4 and 18). A five-by-five construction-reader evaluation also supports choosing construction and answering models separately (Table 19; Figure 7).

  • Reconstruction outcomes and failure analysis (hVT9, wM55). Our responses report grouping statistics, readout comparisons, and an error audit. Appendix H presents the controlled analyses and detailed open-domain failure cases.

  • Construction costs and online updating (we3V, wM55). We report construction token usage by stage (Appendix F.7, Table 17) and clarify that refreshing the final entity view currently requires repeating reconstruction. Buffering, selective reconstruction, compression, and pruning are proposed improvements for future online operation.


We hope these additions clarify ContextFact's contributions, empirical behavior, and scope. Our individual responses to hVT9, we3V, and wM55 provide detailed explanations and supporting results. We welcome any further questions or discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.