acceptodds
Under review as a conference paper at ICLR 2027

When Errors Become Memories: Causal Pathway Tracing in Multi-Turn Memory-Augmented LLMs

Abstract

Long-term memory allows large language models (LLMs) to preserve and reuse information across interaction turns, but it can also transform a localized error into a persistent system-level risk. Existing evaluations of memory-augmented LLMs focus primarily on whether information is correctly stored and retrieved, leaving the causal pathways through which errors persist and re-emerge insufficiently characterized. We introduce a structural causal framework for tracing cross-turn error propagation in memory-augmented LLMs. The framework models user questions, model responses, and memory states as a dynamic causal process, separates memory-update and question-feedback pathways, and uses four counterfactual trajectories to estimate their downstream effects and interaction. Error influence is evaluated through memory retention, natural responses, targeted diagnostic probes, and probability-level error preference. Across models, memory categories, injection positions, and memory mechanisms, error influence generally decreases with interaction distance but remains detectable in internal memory states and controlled behavior after disappearing from natural responses. The memory-update pathway produces more persistent effects on memory-state and probe outcomes, while question-feedback effects can be reactivated by later interactions. Pathway-guided restoration further supports this causal decomposition: repairing the corrupted memory state is more effective than repairing the affected question trajectory alone, and jointly restoring both pathways reduces residual propagation to 0.003, corresponding to a 98.5% reduction relative to the unrepaired condition. These findings show that response-only evaluation can miss latent error traces and that causal pathway tracing provides a principled basis for diagnosing and repairing persistent memory errors. Our code is available at https://anonymous.4open.science/r/Causal-Memory-Pathway-Tracing-33B0.

Primary area:
interpretability and explainable AI
Keywords:
Agent memory, error tracing, causality

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_qb6QOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper evaluates error propagation in long-term memory-augmented LLMs. It identifies two propagation pathways: memory updating and question feedback. For each question, the authors construct closely matched correct and erroneous responses and use them to generate four counterfactual trajectories: fully clean, memory-only contamination, question-only contamination, and joint contamination. Based on this framework, the paper introduces four error measures—memory-state retention, natural-response errors, controlled-probe errors, and probability-level error preference—and quantifies several causal effects, including the total propagation effect (TPE) and pathway-specific effects.

Critical strengths
  • The paper develops a causal model of error propagation in LLMs and constructs four counterfactual trajectories to quantify multiple causal effects.
  • The findings are important such as errors may re-emerge under targeted probing even when absent from natural responses, and generation probabilities may shift toward erroneous answers despite correct final outputs. The observed model-specific susceptibility to different error categories is also informative.
Critical weaknesses

The generalizability of the conclusions is limited for the following reasons:

  • The dataset contains only 144 synthetic multi-turn interactions, with 24 instances per category.
  • The evaluation covers only two 7B models. It remains unclear whether stronger open- or closed-source models, particularly those with self-reflection or verification mechanisms, would exhibit similar memory-error dynamics.
  • Differences between synthetic and real-world interactions, such as user language and behavior, further limit external validity. Moreover, Qwen2.5-32B is used for data generation, while one evaluated response model belongs to the same family, potentially introducing shared model-family biases.

Similarly, although memory repair and joint repair reduce the residual effect by 70% and 98.5%, respectively, these results are obtained entirely from synthetic data and have not been validated in real-world settings.

Questions for authors

See Weaknesses.

4 comments
AuthorsOpenReview ↗

Response to W1: Dataset Size and Additional MemUse Experiments & Response to W2: Larger Models with Reasoning and Reflection

Response to W1: Dataset Size and Additional MemUse Experiments

Thank you for raising the question of dataset size. The 144 cases form the main balanced benchmark, with 24 cases per category. Section 5.3 and Appendix A.3 also report 84 independently constructed additional candidates, bringing the total to 228 candidates and the valid cohorts to 197 Qwen cases and 185 OLMo cases. The expanded cohorts retain the main error-propagation trends.

To extend our evaluation to settings grounded in real conversations, we added experiments using MemUse [1]. We selected 55 candidate cases with sufficient dialogue context and official fact annotations. Target facts were verified against users’ original utterances to construct correct and erroneous responses differing only in the target fact. Ultimately, 50 cases from 19 users yielded validated response pairs and were included in subsequent experiments.

Starting from these real conversation histories and user questions, we ran experiments with Qwen2.5-7B and OLMo3-7B. Each case begins with two shared dialogue turns. At turn 3, we introduce matched correct or erroneous responses, then continue each of the four counterfactual trajectories for 12 additional turns. Across both models, this yields 4,800 post-intervention interaction turns. Subsequent questions are generated dynamically based on the model’s preceding response, rather than being scripted in advance.

At each turn, we assess whether the target error appears in responses and memory states, as well as the model’s probability-level error preference. Detailed results for these experiments are provided in our response to W3.

Response to W2: Larger Models with Reasoning and Reflection

Thank you for this suggestion. We have expanded the evaluation beyond the original two 7B models to larger models and configurations with reasoning and explicit reflection. We added four experimental settings: gpt-oss-120b with medium reasoning effort, and Qwen3-32B in three configurations—non-thinking, thinking, and thinking with explicit reflection.

All four settings use the original 144 candidate cases, the MemoryBank mechanism, error injection at turn 3, and four counterfactual trajectories, followed for the full 12 post-injection turns. Qwen2.5-32B-Instruct is used consistently for subsequent question generation and proposition-level evaluation. The primary analysis includes all candidate cases, including those in which the target error was not successfully written into memory, and uses 5,000 case-level bootstrap resamples. Comparisons across configurations are paired by candidate case.

The additional results show that error propagation is not limited to smaller models: it also occurs in larger models and configurations with reasoning and reflection. Across all four settings, the natural-response total propagation effect (TPE), averaged over the 12 post-injection turns, is positive:

Model and configurationTPE, percentage points [95% CI]
gpt-oss-120b, medium reasoning effort15.74 [11.86, 19.85]
Qwen3-32B, non-thinking9.78 [6.13, 13.54]
Qwen3-32B, thinking14.87 [10.82, 18.98]
Qwen3-32B, thinking with explicit reflection18.11 [14.24, 22.28]

Across all four settings, the natural-response TPE averaged over post-injection turns 2–12 remains positive, with 95% CIs above zero, showing propagation beyond the first turn. Memory-update effects exceed question-feedback effects on memory-state and controlled-probe metrics in all four settings, with all 95% CIs for the differences above zero. With explicit reflection, the natural-response memory-update and question-feedback effects averaged over post-injection turns 1–12 are 14.06 and 7.29 percentage points, respectively. The difference is 6.77 percentage points [95% CI: 2.84, 10.88]. These findings extend the analysis to reasoning and reflection.

Compared with thinking alone, adding explicit reflection reduces the first-attempt error-writing rate from 54.86% to 18.75%, a decrease of 36.11 percentage points [95% CI: 26.39, 45.83]. With up to five attempts, the final valid injection rate decreases from 64.58% to 48.61%, a reduction of 15.97 percentage points [95% CI: 5.56, 26.39]. However, a clear positive propagation effect remains in subsequent interactions. Thus, the tested reflection configuration reduces erroneous memory writes without eliminating downstream propagation, highlighting the need to evaluate these separately and track both the memory-update and question-feedback pathways.

[1] Ryuichi Sumida, Koji Inoue, and Tatsuya Kawahara. 2026. MemUse: Moving Memory Evaluation from Direct QA to Natural Integration in Long-Term Human-AI Conversation. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 18719–18749.

AuthorsOpenReview ↗

Response to W3: Real User Histories and Model Diversity

Thank you for your suggestion. We have added experiments using real user conversation histories and response models from different families to examine whether error propagation also occurs in these settings.

First, we completed experiments based on MemUse, comprising 50 dialogue cases from 19 real users. Each case starts from a real conversation history and its corresponding user question, followed by controlled error injection and 12 turns of simulated interaction. Both Qwen2.5-7B and OLMo-3-7B completed all four counterfactual trajectories for every case. This introduces real user language and historical context while maintaining comparable experimental conditions.

Second, OLMo-3-7B, already included in the paper, belongs to a different model family from Qwen. We have also added gpt-oss-120b from another model family, using medium reasoning effort. Keeping the original question-generation and evaluation settings unchanged, we completed all four counterfactual trajectories for all 144 candidate cases.

The natural-response results from these additional experiments are summarized below:

DatasetResponse modelNumber of casesNatural-response TPE [95% CI]
MemUse real user historiesQwen2.5-7B504.00 [0.30, 9.35]
MemUse real user historiesOLMo-3-7B505.67 [2.54, 9.62]
Original controlled casesgpt-oss-120b14415.74 [11.86, 19.85]

Values represent the total propagation effect averaged over post-injection turns 1–12, expressed in percentage points. They measure the change in the proportion of responses containing the target error between the trajectory with errors introduced through both pathways and the clean trajectory. All reported confidence intervals are entirely above zero.

The MemUse results show that injected errors can influence subsequent responses when interactions begin from real user histories. The OLMo and gpt-oss results further show that this phenomenon also occurs in response models outside the Qwen family. We also compared the effects of the two pathways. The average pathway effects for gpt-oss-120b over post-injection turns 1–12 are reported below:

MetricsMemory-update pathway effect (M)Question-feedback pathway effect (Q)Difference (M − Q) [95% CI]
Natural-response errors7.297.75−0.46 [−4.17, 3.30]
Memory-state errors41.324.6936.63 [28.65, 44.91]
Controlled-probe errors25.693.9921.70 [15.39, 28.01]
Probability-level error preference0.3680.0250.344 [0.099, 0.597]

The first three metrics are reported in percentage points, while the final metric is measured in bits/token. For gpt-oss, the memory-update pathway has a larger effect than the question-feedback pathway on memory-state errors, controlled-probe errors, and probability-level error preference. For example, the controlled-probe effect is 25.69 percentage points for the memory-update pathway, compared with 3.99 percentage points for the question-feedback pathway. This supports the original observation that the memory-update pathway has a greater influence on error retention and subsequent probe responses.

These additional experiments extend the paper’s evidence across both data sources and model families. The MemUse results show that error propagation through the two pathways remains observable when using real users’ conversation histories and questions. The OLMo and gpt-oss results show that this phenomenon also occurs in response models outside the Qwen family. Furthermore, the pathway analysis for gpt-oss supports a larger contribution from the memory-update pathway on memory-state, controlled-probe, and probability-level error preference outcomes. We will include these experiments and their complete results in the revised paper to show how the original findings are supported in settings grounded in real user histories and across different model families.

AuthorsOpenReview ↗

Response to the Concern about Real-World Validation of Repair Results

Thank you for your comment on the real-world applicability of our repair results. To address this concern, we added restoration experiments based on real user histories from MemUse, using 50 dialogue cases from 19 users, and evaluated them on Qwen2.5-7B and OLMo-3-7B. Each case starts from a real dialogue history and user question, undergoes controlled error injection, and then continues with model-generated dialogue. After the end of turn 5, we compare four conditions: No Repair, Question Repair, Memory Repair, and Joint Repair. We then examine performance over turns 6–15 across all 50 cases.

We use the same probe questions to check whether responses under different repair conditions contain the target error and compare them with the clean reference without error injection. A lower MRE indicates that the results after repair are closer to the clean reference.

Response modelRepair methodNo Repair MREMRE after repairMRE reduction [95% CI]
Qwen2.5-7BQuestion Repair0.1300.10618.5% [−11.3%, 46.7%]
Qwen2.5-7BMemory Repair0.1300.05656.9% [19.6%, 84.0%]
Qwen2.5-7BJoint Repair0.1300.04466.2% [26.7%, 86.6%]
OLMo-3-7BQuestion Repair0.1160.09419.0% [−18.2%, 44.6%]
OLMo-3-7BMemory Repair0.1160.09617.2% [−23.3%, 44.9%]
OLMo-3-7BJoint Repair0.1160.07634.5% [12.5%, 54.0%]

All percentage reductions are calculated relative to No Repair. Confidence intervals are computed using 5,000 bootstrap resamples clustered by user.

On Qwen, Memory Repair and Joint Repair reduce MRE by 56.9% and 66.2%, respectively, with confidence intervals supporting improvements in both cases. A direct comparison also shows that Memory Repair outperforms Question Repair. On OLMo, Joint Repair reduces MRE by 34.5%, also supported by the confidence interval.

These additional results show that improvements from repair can also be observed with real user histories, while the benefits vary across models. We will include these experiments and the complete results in the revised manuscript and describe the specific settings for the real user data and model-generated interactions.

Thank you for taking the time to read our paper closely and share your comments. These improvements have strengthened the credibility of our work. We value this exchange and welcome any further questions you may have. We sincerely wish you all the best!

Dear Reviewer,

I hope this message finds you well. As the discussion period is nearing its end, with less than two days remaining, we wanted to ensure that we have addressed all your concerns satisfactorily. If there are any additional points or feedback you would like us to consider, please let us know. Your insights are invaluable to us, and we are eager to address any remaining issues to improve our work.

Thank you for your time and effort in reviewing our paper.

Reviewer_Yu7zOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3(+1)
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a structural causal framework to analyze cross-turn error propagation in memory-augmented LLMs. Using four counterfactual trajectories, it separates memory-update and question-feedback effects and explores pathway-guided error restoration.

Critical strengths

S1: The paper addresses how errors can persist in long-term memory even after disappearing from natural responses, an important issue for reliable memory-augmented agents.

S2: Structured causal framework. The four-trajectory design explicitly distinguishes memory-update and question-feedback pathways, providing a systematic approach to analyzing error propagation.

S3: Multiple complementary metrics. The paper evaluates memory retention, natural responses, controlled probes, and generation preferences, offering a broader view than final-answer accuracy alone.

Critical weaknesses

W1: Although the four trajectories share the same pre-intervention history, it remains unclear whether erroneous responses are retained in the model's dialogue context. Such direct influences could confound the claimed pathway-specific effects.

W2: The validity filter makes the initial memory-state effect of 1.0 true by construction. Moreover, restoration is evaluated on only one answer model and one memory mechanism.

Questions for authors

Q1: How are dialogue history and previous responses handled across the four trajectories to ensure that the intended causal pathways are isolated?

Q2: How sensitive are the results to the validity filter? Can the authors report an analysis over all candidate instances, including failed injections, to clarify the scope of the reported effects?

Q3: Since the reported 98.5% reduction (MRE = 0.003) is attributed to inconsistent evaluator labels for identical outputs, can the authors recompute this result using consistent labels and clarify whether the reported reduction remains valid?

7 comments
AuthorsOpenReview ↗

Response to W1 and Q1: Dialogue History and Pathway Isolation & Response to W2: Validity Filter and Extended Restoration Experiments

Response to W1 and Q1: Dialogue History and Pathway Isolation

Thank you for raising this careful concern. In our experimental setting, the erroneous response is not passed directly to the response model as additional dialogue-history context. Its influence is transmitted through the memory-update and question-feedback pathways.

Following the causal structure in Figure 1, all trajectories start from the same pre-injection history. We independently control whether the correct or erroneous response is supplied to the memory updater and the next-question generator. For trajectories with an error introduced through question feedback, the erroneous response remains in the history used by the question generator. For trajectories with clean question feedback, the corresponding correct response is used instead.

The question generator produces the current question from the recent dialogue history and the previous response. The response model receives only the current question and the current memory state. This input arrangement allows historical information to influence subsequent responses through the mechanisms specified in Figure 1. The pathway effects distinguish where the error enters at the intervention turn; the two pathways are allowed to interact in subsequent turns.

Thus, the erroneous response may remain in the question generator’s history, but it is not supplied as additional context to the response model. Its influence on subsequent questions is included in the question-feedback pathway. The additional direct history-to-response channel raised by the reviewer is therefore not present in the current implementation. We will clarify the inputs to each module in the revised paper.

Response to W2: Validity Filter and Extended Restoration Experiments

Thank you for your suggestion. We have added an all-candidate analysis and extended the restoration experiments to two response models and two memory mechanisms.

Regarding the validity filter, the original filter required the erroneous response to introduce the error into memory successfully, while the correct response did not introduce that error. The initial memory-state effect is therefore 1.0, confirming successful error injection. Whether the error continues to affect subsequent memory states and responses must be assessed using the results from later turns. To examine this, we added four experimental settings: Qwen2.5-7B and OLMo-3-7B, each paired with MemoryBank and Top-3 retrieval memory. Each setting uses the same 144 candidate cases, and cases with failed injections also continue through the subsequent turns and are included in the analysis. After excluding the first post-injection turn, the total propagation effects (TPE) remain positive across all four settings on all four metrics—natural responses, memory states, controlled probes, and answer preference—with all 95% confidence intervals above zero. This shows that subsequent error propagation remains observable across all candidates, including cases with failed injections.

Regarding the restoration experiments, we compared No Repair, Memory Repair, Question Repair, and Joint Repair in all four settings. Restoration was performed before the start of turn 6, followed by evaluation over turns 6–15. The controlled-probe results are as follows:

Response modelMemory mechanismNo Repair MREMemory Repair MREMRE reduction [95% CI]
Qwen2.5-7BMemoryBank0.32360.156251.7% [40.4%, 61.5%]
OLMo-3-7BMemoryBank0.29310.203530.6% [17.9%, 41.0%]
Qwen2.5-7BTop-30.24100.021591.1% [85.3%, 95.8%]
OLMo-3-7BTop-30.45900.022995.0% [91.6%, 97.9%]

MRE measures the mean absolute difference between each condition and the clean trajectory; a lower value indicates that the condition is closer to the clean trajectory. Confidence intervals were computed using 5,000 case-level bootstrap resamples. Memory Repair reduced the controlled-probe MRE in all four settings. Paired comparisons also support its advantage over Question Repair in the average results across these 10 turns. Thus, this restoration finding is supported by two response models and two memory mechanisms. We will include these experiments and the complete results in the revised manuscript.

AuthorsOpenReview ↗

Response to Q2: Sensitivity to the Validity Filter

Thank you for this suggestion. We have added a comparison between the original valid subset and all 144 candidate instances. Including the previously excluded cases reduces the average propagation effects, but the 95% confidence intervals for all four total propagation effects remain above zero for both models.

The original validity filter verifies that the intervention produces the intended difference in memory: the correct response does not introduce the target error into memory, while the erroneous response successfully does so. This allows us to study how an error propagates after it has been written into memory.

The additional analysis retains the original 125 valid Qwen cases and 120 valid OLMo cases. We conducted supplementary runs for the 19 Qwen cases and 24 OLMo cases that originally failed the validity check, producing complete four-trajectory results for all 144 candidates per model. Because the original failed-run records did not preserve the final injection state, these supplementary runs reused the pre-injection histories and questions, repeated the injection procedure, and tracked the following 12 turns. Group assignments in the analysis remain based on the original validity records.

The table below compares the total propagation effect (TPE) averaged over post-injection turns 1–12:

ModelMetricsOriginal valid subset: TPE [95% CI]All 144 candidates: TPE [95% CI]Relative decrease
Qwen2.5-7BNatural-response errors18.60 [14.13, 23.40]16.96 [12.73, 21.24]8.8%
Qwen2.5-7BMemory-state errors37.53 [31.93, 43.07]33.22 [27.89, 38.83]11.5%
Qwen2.5-7BControlled-probe errors35.20 [29.93, 40.40]32.12 [27.02, 37.21]8.8%
Qwen2.5-7BProbability-level error preference0.898 [0.738, 1.052]0.824 [0.680, 0.972]8.3%
OLMo-3-7BNatural-response errors22.15 [17.36, 27.08]21.59 [17.30, 26.10]2.6%
OLMo-3-7BMemory-state errors30.28 [25.83, 34.86]28.07 [23.84, 32.47]7.3%
OLMo-3-7BControlled-probe errors26.11 [21.67, 30.49]24.94 [20.83, 29.22]4.5%
OLMo-3-7BProbability-level error preference0.527 [0.416, 0.649]0.465 [0.357, 0.574]11.7%

The first three metrics are reported in percentage points, while probability-level error preference is measured in bits/token. Confidence intervals are calculated using 5,000 bootstrap resamples, with complete cases as the resampling unit.

These results indicate that the estimated magnitude of propagation depends on the analyzed cohort, but overall error propagation remains evident after including the previously excluded cases. Compared with the original valid subset, total propagation effects decrease by approximately 8.3%–11.5% across the four outcomes for Qwen and 2.6%–11.7% for OLMo. For example, the natural-response TPE decreases from 18.60 to 16.96 percentage points for Qwen and from 22.15 to 21.59 percentage points for OLMo, with the confidence intervals remaining entirely above zero.

We also reassessed the two pathways. Averaged over these 12 turns, the memory-state, controlled-probe, and probability-level error preference outcomes continue to support a larger memory-update pathway effect than question-feedback pathway effect for both models. For natural responses, however, there is no clear statistical evidence of a difference between the two pathways. We will therefore report the original valid-subset analysis alongside the supplementary all-candidate analysis in the revised paper. We will clarify that the former characterizes propagation after the target error has been successfully written into memory, while the latter also includes candidates that originally failed the validity check, and report the pathway comparisons separately for each outcome.

AuthorsOpenReview ↗

Response to Q3: Recalculation with Consistent Labels

Thank you for this suggestion. We have applied consistent evaluator labels to identical outputs in the original Table 1 restoration experiment and recomputed the results. Joint Repair’s MRE changes from 0.003 to 0, and its reduction relative to No Repair changes from 98.5% to 100%.

Specifically, across the 730 post-restoration records from the 73 test cases in the original Table 1 experiment, the probe responses under Joint Repair are identical to those of the clean trajectory. However, the evaluator assigned different labels to two of these identical response pairs, producing the previously reported nonzero residual. We now assign the clean-reference label to any output that exactly matches the clean reference for the same target error, applying this rule consistently across all repair methods.

The same consistency rule also resolves one Memory Repair record whose probe response matches the clean reference but received a different label. Its MRE changes from 0.0534 (rounded to 0.053 in the original manuscript) to 0.0521, and its reduction relative to No Repair changes from 70.0% to 70.8%. Question Repair remains at an MRE of 0.1288 and a reduction of 27.7%. The recalculated Memory Repair minus Question Repair difference is −0.0767 (95% CI: [−0.1260, −0.0301]), continuing to support a greater improvement from Memory Repair on controlled probes. Joint Repair has an MRE of 0, consistent with its role as a clean-state restoration reference. We will update these values and clarify this reference role in the revised paper.

We would be happy to clarify any remaining questions. Thank you for your careful review and valuable suggestions, and we wish you all the best!

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

This paper studies how errors spread across multiple turns in memory-augmented LLMs. The authors identify two main propagation pathways: memory updates and question feedback. They use four counterfactual trajectories to examine how errors persist in memory, affect model responses, and influence subsequent predictions. Experiments with two LLMs, six memory categories, different error injection positions, and two memory mechanisms show that errors can remain in memory even when they no longer appear in model responses. The memory-update pathway also tends to have a longer-lasting effect. Finally, the authors show that correcting memory is more effective than correcting question trajectories alone, while correcting both reduces residual error propagation by 98.5%.

Critical strengths
  1. This paper focuses on a meaningful research problem. They investigate how memory errors propagate over time and persist even after becoming invisible.

  2. The paper presents a rigorous causal design. They introduce a four-trajectory counterfactual framework to disentangle the effects of memory, question and their interactions on error propagation.

  3. They evaluate errors at four levels: memory state, natural answer, controlled probe and probability preference.

Critical weaknesses
  1. There are some overlaps between this paper and existing studies, such as MemTrace[1] and “How Memory Management Impacts LLM Agents: An Empirical Study of Experience-Following Behavior“.

  2. This paper mainly comfirms an existing observation but does not propose a new and reliable sollution.

  3. The 98.5% reduction in error propagation is not surprising because the restoration experiments rely on access to correct trajectories. This is essentially an oracle setting.

  4. The experiments rely entirely on synthetic data, with no validation on real agent envrioments.

[1] MEMTRACE: TRACING AND ATTRIBUTING ERRORS IN LARGE LANGUAGE MODEL MEMORY SYSTEMS

Questions for authors

Same as Weakness part

6 comments
AuthorsOpenReview ↗

Response to W1: Comparison with Related Work

Thank you for highlighting these two closely related studies. Our specific contribution is to use four counterfactual trajectories for the same case to quantify the downstream effects of an erroneous response entering through memory updating and question feedback, including their interaction. Below, we explain how this analysis complements MemTrace’s operation-level error attribution and Xiong et al.’s study of memory management.

1.Comparison with MemTrace [1]. MemTrace represents memory-system execution through memory evolution graphs. It traces information flow across operations to identify the operations responsible for failed responses and uses these attribution results to guide prompt optimization. Its main focus is identifying where failures occur within a memory system and using this information to improve the system. Our work examines how the same erroneous response affects subsequent interactions when introduced through different pathways. Within a structural causal model, we distinguish the memory-update pathway from the question-feedback pathway. By independently controlling the inputs to these two mechanisms, we construct four counterfactual trajectories and estimate their downstream effects and interaction across turns. Complementing MemTrace’s operation-level failure attribution, our framework provides a pathway-level causal decomposition of cross-turn error propagation. This allows us to compare the contributions of internal memory and subsequent questions to persistent error influence.

2.Comparison with “How Memory Management Impacts LLM Agents: An Empirical Study of Experience-Following Behavior”[2]. Xiong et al. study how memory addition and deletion affect agents’ experience-following behavior. They show that errors in past experiences can accumulate and impair performance on subsequent tasks. Their work establishes the presence of error propagation in agent memory and highlights the importance of memory management and experience quality. Building on this understanding, our work further distinguishes the downstream effects of errors introduced through memory updating and question feedback. Specifically, we intervene on these two error-entry mechanisms within the same set of cases, allowing us to quantify each pathway’s contribution and their interaction. This adds a pathway-level explanation to analyses of overall performance under different memory management strategies.

We also measure error influence through four complementary outcomes: natural responses, memory states, targeted diagnostic probes, and probability-level error preference. These outcomes distinguish whether an error remains in memory, whether it re-emerges when directly queried, and whether it continues to bias generation probabilities toward the erroneous answer. This allows us to identify latent error influence that is no longer visible in natural responses. Section 5.5 further compares Memory Repair and Question Repair to examine whether pathway effects can guide the choice of intervention location.

We will expand the Related Work section with these comparisons to clarify both the connections to prior studies and our specific contributions in causal decomposition of the two pathways, measurement of cross-turn error propagation, and pathway-guided state restoration.

[1] Deng, X., et al. (2026). MemTrace: Tracing and Attributing Errors in Large Language Model Memory Systems.

[2] Xiong, Z., et al. (2025). How Memory Management Impacts LLM Agents: An Empirical Study of Experience-Following Behavior.

AuthorsOpenReview ↗

Response to W2: Causal Interpretability and Practical Value

Thank you for your comment. We will clarify the potential uses of our framework and the experimental evidence supporting them. Our core contribution is a causal framework for explaining, diagnosing, and evaluating cross-turn error propagation in memory-augmented LLMs, consistent with the focus of the Interpretability and Explainable AI track. We model the joint evolution of user questions, model responses, and memory states within a unified structural causal model. Four counterfactual trajectories allow us to quantify the memory-update pathway effect, the question-feedback pathway effect, and their interaction. This framework offers three potential uses.

1.Assessing whether error influence has actually subsided. An error disappearing from natural responses does not necessarily mean that its influence has disappeared. We track the same target error through four complementary outcomes: natural responses, memory states, targeted diagnostic probes, and probability-level error preference. Our experiments show that an error may remain in memory after disappearing from natural responses, re-emerge when directly queried, or continue to bias generation probabilities toward the erroneous answer. The framework therefore helps detect latent error traces that natural-response evaluation can miss. It also provides a more complete basis for assessing whether an intervention reduces error influence over subsequent turns.

2.Providing evidence for choosing where to intervene. Separating the two propagation pathways allows us to compare the effects of restoring memory states and restoring error-affected dialogue trajectories. Section 5.5 of the original manuscript reports pathway-guided state restoration using Qwen2.5-7B-Instruct with the MemoryBank-style mechanism. Memory Repair and Question Repair reduce controlled-probe MRE by 70.0% and 27.7%, respectively. The paired comparison supports a greater reduction from Memory Repair. Furthermore, question-feedback pathway effects are positively correlated with the benefits of Question Repair (ρ = 0.335, 95% CI [0.158, 0.493]). These results support a practical use of pathway analysis: identifying which propagation mechanism warrants closer inspection and informing comparisons between intervention locations. The appropriate intervention should depend on the outcome of interest and the conditions of the case, rather than assuming that one pathway is always more important.

3.Supporting the evaluation and improvement of memory mechanisms. We also use the same framework to compare continuously updated MemoryBank with episodic retrieval. The experiments reveal different propagation patterns: error influence can decay as memory is updated, while erroneous entries may persist in episodic storage. Retaining an error in memory does not necessarily mean that it will immediately appear in a response. Developers can therefore examine how a memory mechanism retains errors, when those errors are reused, and how long their influence persists. This provides a basis for evaluating memory updating, retrieval, and deletion strategies.

Experiments across models, memory categories, memory mechanisms, and independent evaluators, together with supplementary analyses under different case-selection criteria, provide supporting evidence for this propagation analysis. The current state-restoration experiments involve known target errors and use clean reference states for controlled comparisons. Their results provide an experimental foundation for developing automated diagnosis and repair methods. We will make these uses, supporting evidence, and conditions of applicability clearer in the revised paper, emphasizing how explanations of propagation mechanisms can inform testable approaches to diagnosis and intervention.

AuthorsOpenReview ↗

Response to W3: Oracle Setting and the Role of State Restoration

Thank you for your comment. We would like to clarify that our main comparison is between two partial restorations applied to the same error instance: restoring only the memory state and restoring only the error-affected dialogue state. We examine which intervention more effectively reduces downstream error influence. All restoration conditions use the clean trajectory of the same case as a reference, allowing us to compare memory restoration and question-feedback state restoration under matched conditions. Joint Repair provides a complete-state restoration reference, while the comparison between Memory Repair and Question Repair tests how much improvement each repair target provides.

The original manuscript reports that Memory Repair and Question Repair reduce controlled-probe MRE by 70.0% and 27.7%, respectively, relative to No Repair. Memory Repair has a lower MRE, with a paired Memory Repair minus Question Repair difference of −0.075 (95% CI: [−0.125, −0.030]). Thus, in this experimental setting, restoring memory is more effective than restoring only the dialogue state, consistent with our pathway analysis.

The value of this experiment is therefore in testing whether intervening at the location suggested by pathway analysis produces a greater improvement. We will make these comparisons clearer in the revised paper. These results provide experimental evidence to inform the choice of intervention location and may help guide the development of automated repair methods.

Reviewer_1D4nOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies how a single wrong answer spreads through later turns in a chatbot with long-term memory. The authors describe two ways an error can spread: it gets saved into memory, or it shapes the user’s next question. To separate the two, they plant a wrong answer at one turn and run four versions of the conversation, letting the error into memory, into the next question, both, or neither. They then track the error over 12 later turns by checking the memory, the model’s normal answers, direct test questions, and the model’s preference for the wrong answer. Errors fade over time but often stay in memory or show up under direct questioning after they disappear from normal answers. Errors saved in memory last longer, and fixing the memory removes more of the error than fixing the conversation.

Critical strengths
  1. Comprehensive evaluations that measures errors at several levels: Checking the memory, normal answers, direct test questions, and answer probabilities shows that an error can disappear from normal answers but still be stored or easy to bring back. This is a useful point for anyone evaluating these systems.
  2. Multiple robustness checks: The authors test different random seeds, answer lengths, injection turns, and outside judge models. They report invalid cases openly and release their code.
Critical weaknesses
  1. Potential confirmation bias? Only cases where the wrong answer was successfully saved to memory are kept, so the memory effect at the first turn is 1.0 by design. There is no matching check that the error actually reached the next question. This makes the finding that the memory path is stronger partly built into the setup.
  2. Joint repair just resets the conversation back to the clean version, so of course almost no error remains. The authors even say the small leftover (0.003) is judge noise. Highlighting the 98.5% reduction in the abstract makes the result sound bigger than it is. Also, every repair method assumes you already know where the error is, which isn’t true in real use.
  3. Earlier work has already shown that errors saved in memory keep affecting later answers (e.g., Xiong et al., 2025; MemTrace; studies on memory poisoning and error snowballing). It’s also expected that errors fade over time and that saved entries stay in episodic memory.
Questions for authors
  1. Can you add recommendations to fix memory aided errors in real/practical systems?
  2. Do the results hold with a judge model that wasn’t also used to build the test cases?
  3. How would the results change if you also required the error to show up in the next question, not just in memory, when choosing valid cases?
4 comments
AuthorsOpenReview ↗

Response to W1: Potential confirmation bias?& Response to W2: Joint Repair and Practical Applicability

Response to W1: Potential confirmation bias?

Thank you for pointing out that the validity filter may affect the pathway comparison. To address this concern, we included all cases that had originally failed the memory-write validity check, including those in which the error was not successfully written into memory, and generated complete subsequent trajectories for these cases. We retained them regardless of whether the error was successfully written into memory during the rerun, yielding all 144 candidate cases for each model.

The original filter was used to study error propagation after successful memory writing. By extending the analysis to all candidates, we examined how this filter affects the effect sizes and pathway comparison. In the all-candidate analysis, cases were retained even if the error was not successfully written into memory or carried into the next question. Across all candidates, the average effects of the two pathways over the 12 complete dialogue turns following error injection are as follows:

ModelMetricMemory-update pathway effectQuestion-feedback pathway effectDifference (memory − question) [95% CI]
QwenNatural responses0.0980.0950.003 [−0.040, 0.046]
QwenMemory states0.2910.0950.195 [0.134, 0.256]
QwenControlled probes0.2490.0760.172 [0.119, 0.225]
QwenProbability preference0.7610.2000.561 [0.408, 0.725]
OLMoNatural responses0.1130.1060.008 [−0.039, 0.053]
OLMoMemory states0.2110.1010.110 [0.068, 0.150]
OLMoControlled probes0.1400.0870.053 [0.012, 0.094]
OLMoProbability preference0.2910.0950.196 [0.105, 0.289]

Confidence intervals were computed using 5,000 case-level bootstrap resamples. The first three metrics are differences in error rates on a 0–1 scale; for example, 0.10 represents 10 percentage points. Probability preference is measured in bits/token. As shown in the table, when averaged over all cases, the memory-update pathway effect remains larger than the question-feedback pathway effect for both models on natural responses, memory states, controlled probes, and probability preference. For the latter three metrics, the 95% confidence intervals for the differences are all above zero. Overall, when all cases are retained, the average advantage of the memory-update pathway remains clear across these metrics for both models, consistent with the trends in the original manuscript but supported by more convincing evidence. We will use these all-candidate results in the revised manuscript. Thank you again for your thoughtful suggestion!

Response to W2: Joint Repair and Practical Applicability

Thank you for this suggestion. Our error restoration experiments aim to answer the following question: once an error has been located, is repairing memory or the question-side state more effective at reducing subsequent propagation? We use the same clean trajectory as the restoration reference to compare the effects of repairing different components. Accordingly, Joint Repair serves as the reference condition in which the clean state is fully restored. Specifically, Memory Repair restores memory while retaining the dialogue history affected by the error, whereas Question Repair restores the question-side history and the previous response while retaining the contaminated memory.

Table 1 in the original manuscript reports that, in the restoration experiment using Qwen2.5-7B and MemoryBank, the controlled-probe MRE values for No Repair, Question Repair, and Memory Repair are 0.178, 0.129, and 0.053, respectively. The latter two reduce the residual effect by 27.7% and 70.0%, respectively. This comparison shows that, under the current setting, repairing memory removes more of the subsequent effects, which is the main finding of the restoration experiments. The abstract of the revised manuscript will emphasize the comparison between Memory Repair and Question Repair, and the main text will explain the role of Joint Repair as a reference for complete state restoration.

These experiments use known errors and clean reference states to provide controlled evidence for selecting a repair strategy after an error has been confirmed. In practical systems, erroneous facts can first be confirmed through user corrections or verification against reliable sources, followed by an examination of the relevant memory and subsequent questions. Please see the following Q1 for specific practical recommendations.

AuthorsOpenReview ↗

Response to W3: Comparison with Prior Work

Thank you for highlighting these related studies. Previous work has shown that errors in memory and previously generated errors can affect subsequent responses. Building on these findings, our work further distinguishes and quantifies the memory-update and question-feedback pathways. Specifically:

  1. Xiong et al. [1] study how memory management affects experience-following behavior, including how erroneous experiences influence subsequent task performance. Our work further examines how much subsequent influence the same error produces through memory updates and question feedback, quantifying the relative contributions of the two pathways and their interaction.

  2. MemTrace [2] studies how failures in memory systems can be traced to specific operations and uses the attribution results to guide system improvements. Our work further examines how errors continue to propagate through memory updates and question feedback in subsequent interactions: starting from the same pre-intervention state, we use four counterfactual trajectories to separately estimate the effects of the two pathways and their interaction.

  3. Memory poisoning research [3] studies how contaminated information enters memory and influences subsequent agent behavior. Our work further examines how the subsequent effects differ when the same error enters the memory updater, the next-question generator, or both. We also compare Memory Repair and Question Repair to test how restoring different states affects subsequent propagation.

  4. Error snowballing research [4] studies how earlier erroneous responses lead to further errors in subsequent generation. Our work further examines whether an error continues to affect the system after it is no longer visible in natural responses. We jointly evaluate natural responses, memory states, controlled probes, and probability preference to distinguish “errors that remain stored” from “errors that still affect responses or answer preferences.”

Regarding the reviewer’s observations about error decay and persistent retention in episodic memory, this distinction allows us to further determine whether errors retained in memory can still be elicited through controlled probing or continue to shift the model’s answer preferences. Using the same controlled-probe questions also allows us to compare residual effects across trajectories under consistent questioning conditions. We will add these point-by-point comparisons to the revised manuscript to clarify the connections and differences between our work and previous studies.

[1] Zidi Xiong, Yuping Lin, Wenya Xie, Pengfei He, Zirui Liu, Jiliang Tang, Himabindu Lakkaraju, and Zhen Xiang. 2026. How Memory Management Impacts LLM Agents: An Empirical Study of Experience-Following Behavior. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 623–645.

[2] Xinle Deng et al. 2026. MemTrace: Tracing and Attributing Errors in Large Language Model Memory Systems. arXiv preprint arXiv:2605.28732.

[3] Pritam Dash, Tongyu Ge, Aditi Jain, Tanmay Shah, and Zhiwei Shang. 2026. From Untrusted Input to Trusted Memory: A Systematic Study of Memory Poisoning Attacks in LLM Agents. In Second Workshop on Agents in the Wild: Safety, Security, and Beyond (AIWILD) at ICML 2026.

[4] Muru Zhang, Ofir Press, William Merrill, Alisa Liu, and Noah A. Smith. 2024. How Language Model Hallucinations Can Snowball. In Proceedings of the 41st International Conference on Machine Learning, PMLR 235, pages 59670–59684.

AuthorsOpenReview ↗

Response to Q1: Practical Guidance for Error Diagnosis and Repair & Response to Q2: Independent-Judge Validation

Response to Q1: Practical Guidance for Error Diagnosis and Repair

Thank you for your suggestion. Our results provide experimental evidence for selecting repair targets and verifying repair effectiveness after an error has been confirmed. In the controlled experiments, we use known target errors and clean reference states so that the effects of repairing different targets can be compared under the same conditions. The original manuscript reports that, in the Qwen2.5-7B experiment with MemoryBank, Memory Repair and Question Repair reduce the controlled-probe MRE by 70.0% and 27.7%, respectively.

Based on this result, we recommend that, after confirming an error through user correction or a reliable source, systems in practice check whether the target error remains stored in memory and use controlled probes to determine whether it continues to influence responses. If residual influence remains, the relevant memories can be prioritized for verification, while subsequent questions should also be checked for reliance on the erroneous premise.

For cases where a factual error has been confirmed but the specific corrupted memory entry remains unidentified, the controlled-intervention approach in this paper can be further applied: keep the same set of probe questions fixed, generate responses using the full memory and with an individual memory entry temporarily masked, and compare changes in the controlled-probe outcome $Y_C$ and probability-level error preference YPY_P. If masking an entry causes the error to be elicited less frequently or clearly reduces preference for the erroneous answer, prioritize verification of that entry’s source and correct it. In future work, we will further explore this local memory-masking approach, using changes before and after intervention as clues for prioritizing verification of relevant entries.

After repair, we recommend checking both the memory-state outcome YMY_M and the controlled-probe outcome YCY_C. This paper finds that, after an error disappears from natural responses, it may still remain in memory or be elicited again through targeted diagnostic probing. Therefore, systems in practice should check whether the target error is still present in memory and whether it still influences responses concerning that fact. If residual influence remains, the relevant memories and subsequent questions that depend on the erroneous premise should be examined again.

These recommendations connect error verification, inspection of potentially corrupted memories, and verification of repair effectiveness: user feedback or external verification provides the correct fact, while quantitative changes before and after intervention help determine priorities for inspection and repair. We will include local memory masking in the revised manuscript as a practical recommendation based on the experimental results.

Response to Q2: Independent-Judge Validation

We thank the reviewer for raising the possibility of dependence between test-case construction and the evaluation model. We have reported an independent-judge evaluation in Section 5.2 and Appendix A.5. Qwen2.5-32B-Instruct participated in constructing the test cases; we then used GLM-5.2, which was not involved in their construction, to reassess all 35,280 proposition-level outcomes, and GPT-5.5 to provide an additional assessment of a stratified sample of 900 outcomes. Agreement between the Qwen and GLM judgments was 91.9% (Cohen’s κ = 0.709). Under the complete GLM reassessment, the direction of all 72 reported total propagation-effect estimates remained unchanged, and the temporal effect curves correlated with the original results at 0.975–0.996. These findings indicate that the choice of evaluation model does not alter our main conclusions about the temporal patterns and pathways of error propagation. We appreciate the reviewer’s suggestion and will make this existing validation and its results more prominent in the main text.

AuthorsOpenReview ↗

Response to Q3: Sensitivity to Next-Question Filtering

Thank you for this suggestion. Among the cases that passed the original memory-write validity check, we further retained those whose next question in the jointly contaminated trajectory (ee) was judged by the original evaluator to contain the target error, and recomputed the pathway effects. This condition retained 81 of 125 Qwen cases and 66 of 120 OLMo cases. All four counterfactual trajectories were retained in full for each selected case, allowing the memory-update and question-feedback pathways to be compared on the same set of cases.

Within the 12-turn post-injection observation window used in the original paper, the total propagation effects for all four metrics remain positive for both models, with all 95% confidence intervals above zero. This indicates that error propagation persists after adding the next-question selection criterion. The average effects of the two pathways are reported below:

ModelMetricsMemory-update pathway effect (M)Question-feedback pathway effect (Q)Pathway difference (M − Q) [95% CI]
QwenNatural response0.12350.1399−0.0165 [−0.0741, 0.0381]
QwenMemory state0.32920.15840.1708 [0.0864, 0.2552]
QwenControlled probe0.28400.14200.1420 [0.0658, 0.2171]
QwenProbability-level error preference0.77240.21930.5531 [0.3293, 0.7711]
OLMoNatural response0.11620.1641−0.0480 [−0.1275, 0.0303]
OLMoMemory state0.24240.14520.0972 [0.0253, 0.1692]
OLMoControlled probe0.17420.14270.0316 [−0.0278, 0.0896]
OLMoProbability-level error preference0.34650.18450.1621 [0.0439, 0.2836]

All results are mean values. The pathway difference is the memory-update pathway effect minus the question-feedback pathway effect. The first three metrics are expressed as probability differences, while probability-level error preference is measured in bits/token.

Under this condition, the memory-state and probability-level error preference metrics still support a larger memory-update pathway effect for both models. Meanwhile, the question-feedback pathway effect increases: for natural responses, it rises from 0.1000 to 0.1399 for Qwen and from 0.1125 to 0.1641 for OLMo. The 95% confidence intervals for these increases are [0.0067, 0.0745] and [0.0124, 0.0939], respectively.

We will add this conditional analysis to show that error propagation persists after introducing the next-question criterion and report the relative effects of the two pathways for each metric.

Thank you for the time and care you have devoted to our paper. Your comments helped us strengthen the evidence and explain our findings more clearly, and we sincerely appreciate the thought behind them. We would be happy to discuss any remaining questions. We sincerely wish you all the best !

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Responses

We sincerely thank all reviewers for their careful review and valuable suggestions. In response, we conducted additional experiments and addressed ALL the comments point by point. Our main responses are summarized below:

1. Causal pathways and the validity filter (Reviewers 1D4n and Yu7z): we clarified that the response model receives only the current question and memory, while dialogue history influences subsequent interactions through the question generator. We added an analysis of all 144 candidate cases per model, retaining cases that failed the validity check, and further analyzed cases where the next question contains the target error. After including all candidates, the natural-response total propagation effects for Qwen and OLMo are 16.96 and 21.59 percentage points, respectively, with both 95% confidence intervals entirely above zero. Error propagation remains observable, while the relative contributions of the two pathways vary with the selection criteria and evaluation metrics.

2. Novelty and the significance of the restoration experiments (Reviewers 1D4n and eoun): to the best of our knowledge, this is the first study to use a structural causal model and four counterfactual trajectories within a unified causal framework to separately intervene on the memory-update and question-feedback pathways and quantify their contributions to cross-turn error propagation and their interaction. We clarified the role of Joint Repair as a reference for complete state restoration and explained how the comparison between Memory Repair and Question Repair provides evidence for selecting intervention targets. In Table 1 of the original manuscript, these two methods reduce controlled-probe MRE by 70.0% and 27.7%, respectively.

3. Generalizability across data sources and models (Reviewers eoun and qb6Q): we added 50 cases from 19 users in MemUse, using real conversation histories and user questions, followed by controlled error injection and model-generated interactions. The natural-response total propagation effects for Qwen and OLMo are 4.00 and 5.67 percentage points, respectively, with both 95% confidence intervals entirely above zero. We also added gpt-oss-120b and different configurations of Qwen3-32B, including reasoning and explicit reflection. The natural-response total propagation effects are positive in all four settings. Specifically, the effect for gpt-oss-120b is 15.74 percentage points. For Qwen3-32B, compared with thinking alone, adding explicit reflection reduces the first-attempt error-writing rate from 54.86% to 18.75%, while downstream propagation remains observable.

4. The reliability of restoration and evaluation (Reviewers 1D4n, Yu7z, and qb6Q): we extended the restoration experiments to two response models and two memory mechanisms. Across the four settings, Memory Repair reduces controlled-probe MRE by 30.6%–95.0%, with paired comparisons supporting its advantage over Question Repair in every setting. In the additional MemUse restoration experiments, Memory Repair and Joint Repair reduce MRE by 56.9% and 66.2%, respectively, for Qwen, while Joint Repair reduces MRE by 34.5% for OLMo. All these improvements are supported by the confidence intervals. We also clarified the independent-judge validation already reported in the paper and recomputed the restoration metrics using consistent labels for identical outputs.

The above summarizes our point-by-point responses and the main supplementary evidence. We sincerely thank the reviewers again for their specific suggestions, which have helped us strengthen the experimental evidence and present our findings more clearly.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.