acceptodds
Under review as a conference paper at ICLR 2027

MemResidue: Measuring Risk Propagation in Evolving Agent Memory

Abstract

Persistent memory reuses experience across tasks, creating a channel for risk to travel across memory updates. Safety evaluation of task outcomes can miss this process because a safe current result can coexist with contaminated memory that threatens later tasks. We call the memory left by risk propagation memory residue. The MemResidue Test tracks this residue to measure the extent of propagation across memory evolution. Its paired protocol follows three rounds of memory evolution in Answer and Execution tracks. Across 18 models and 191 cases, residue forms in every tested model, averaging 65.1% in Answer and 32.6% in Execution. Paired interventions show that retaining prior risky memory increases residue in the next update and can increase unsafe outcomes. Generic SafePrompt reduces residue formation. Surviving residue remains highly retrievable, and the phenomenon persists in a representative general memory architecture and under a dedicated memory safety mechanism. Risk propagation in evolving agent memory therefore constitutes an urgent safety dimension for agents.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM agents, agent memory, memory safety, risk propagation, safety evaluation, persistent memory

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_5W79OpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies how dangerous information in agent memory can be retained and propagated across tasks, with a focus on whether the safety of current behavior is sufficient to indicate that newly written memories are also safe. The authors design a three-stage evaluation pipeline: the agent first encounters a harmful source and generates memories, the original source is then hidden, and paired experiments with memory retention or deletion are conducted to measure the risks preserved in subsequent memories and downstream behaviors. Experiments across 18 models and 191 question-answering and tool-use cases show that retaining first-generation memories increases the proportion of residual risks in second-generation memories and significantly increases unsafe behaviors in question-answering scenarios. The study also identifies cases where behaviors are judged safe while the corresponding memories are still recognized as retaining potential risks. The authors further validate these measurements through human review, evaluator replacement, repeated runs, and memory-content interventions, showing that safety prompts and memory filtering can mitigate some risks. The main contribution of this work is a unified evaluation framework that jointly examines memory contents, subsequent retrieval, and actual agent behaviors to analyze how harmful information persists during the memory update process.

Critical strengths

The paper identifies an important yet easily overlooked safety issue for the community: safe current behavior does not necessarily imply that risks have been eliminated from memory. The paper studies how harmful information written into memory can continue to propagate into subsequent memories after the original source has been removed, and investigates this process through paired experiments that retain or delete newly generated memories. Evaluating only current responses or tool executions may overlook harmful information that the system continues to store and transmit. This work reveals an important evaluation blind spot and provides concrete evidence for moving beyond behavior-level safety checks toward assessing the safety of memory update processes.

The paper provides comprehensive and mutually supporting empirical evidence for the core phenomenon of risk propagation. Across 18 models, retaining first-generation memories increases the average residual risk in second-generation memories by 12.2% and 5.8% in question-answering and tool-use scenarios, respectively, with positive propagation gaps observed for 16 and 13 models in the two settings. Human evaluations also recover positive propagation differences in both scenarios, while additional experiments adapting structured memory with A-MEM observe the same trend. The consistency across models, human judgments, and different memory organization mechanisms provides stronger support for the paper's central claim that risks can propagate through memory updates.

Critical weaknesses

The research questions are clearly defined, and the experimental validation is relatively comprehensive. I did not identify any major issues that would affect the core conclusions. Reproducibility of the complete memory evolution process could still be further strengthened. The current repeated experiments fix the parent memories and mainly examine the stability of subsequent propagation under given memory conditions, but do not cover the impact of variations in initial memory generation on the entire evolution process. Adding a small number of independent repeated runs starting from the first task would help verify whether the formation and propagation of risks remain consistent across complete executions, further strengthening the reliability of the conclusions.

Questions for authors

Could the authors provide a small number of independent repeated runs starting from the first task, with each generation of memories regenerated, to examine whether the main findings on risk formation and propagation remain consistent across complete executions? This would complement the current stability validation with fixed parent memories and further strengthen the reliability of the conclusions.

3 comments
AuthorsOpenReview ↗

Response to 5W79: Full-History Repeatability Across 18 Models

Thank you for highlighting the importance of testing variation across the full memory-generation process, beyond our earlier fixed-parent repeats.

The added full-history repetitions show that the aggregate finding persists when the memories themselves are regenerated. Fixed-parent repeats test downstream variation conditional on a given memory, leaving open whether the overall finding depends on a particular initial generation. We therefore repeat the complete history from Task 1 five times on all 18 models and 191 cases. With freshly generated G1 and G2 in every run, the panel-level Formation and Propagation results remain close to the original findings, and mean Propagation is positive in both tracks in all five runs.

Each repetition starts from Task 1 and carries its newly generated memories through the subsequent tasks. We retain the original initial memory pools, all 108 Answer and 83 Execution cases, model settings, top-3 retrieval, task budgets and scoring rules. Each run starts Task 1 afresh to generate G1. Task 2 regenerates both responses and both G2 branches under the within-run G1-kept/removed comparison. Task 3 regenerates both responses using that run's G1 and natural kept-G2 for the G2-available/unavailable comparison. Generated memories are shared where required for within-run pairing, but never reused across runs. All five repetitions are included: 17,190 planned histories, with 17,140 complete through Task 3.

The five-run averages remain close to the original main-table results, with modest variation between panel-level runs. Formation measures strict residue in G1; Propagation is the kept-minus-removed strict-residue rate in G2. The table reports the mean ± sample SD across five run-level, equal-weight 18-model averages.

TrackFormation (%)Propagation (pp)
Answer64.79 ± 0.8511.94 ± 0.69
Execution32.27 ± 1.296.11 ± 1.23

The original 18-model references are 65.12% / 32.62% for Formation and 12.19 / 5.76 pp for Propagation (Answer/Execution). Across the five new runs, Propagation is 12.79, 12.12, 11.03, 11.50, 12.29 pp for Answer and 4.54, 5.28, 7.65, 6.85, 6.22 pp for Execution.

These repetitions cover variability introduced at initial memory generation and carried through subsequent evolution. These run SDs describe panel averages; median within-model Propagation run SDs are 2.71/3.47 pp (Answer/Execution). For each metric, we use model–case units evaluable in all five runs, average cases within each model/run, and weight models equally. Missing scores are excluded.

We have incorporated these results into the main text and provided the full supporting tables and per-case scores:

  • Main text, §5.1: the added full-history repeatability paragraph reports the design, key means and run SDs, with the original main-table references.
  • Appendix P, “Full-history repeatability”: a dedicated addition gives the regeneration and pairing protocol, overview figure, all five results for each model in separate Answer/Execution tables, auxiliary readouts, bootstrap intervals and coverage. A final case-study section compares memory content across repeated histories for three cases.
  • Supplementary Code/Data, revision_addendum/E01_full_history/: tables/case_scores.csv provides model–case–run scores; tables/by_run.csv, tables/by_model.csv and tables/summary.csv provide the summaries. The module includes the case manifest, scoring rules, analysis configuration and analysis.py for recomputation. case_studies/ supplies task context and all five G1/G2-kept/G2-removed text comparisons for each selected case.

Taken together, this validation covers all 18 models and the 191-case benchmark used for the original main table, with five independent repetitions of the full generation-and-evolution protocol. The aggregate Formation and Propagation results remain close to the original findings, and both tracks retain positive mean Propagation in every run. This full-panel replication directly addresses the fixed-parent concern: the central memory-state finding persists across independently regenerated histories, rather than depending on one fixed set of generated parents.

Thank you for your response, you’ve cleared up my main concerns. I will keep my positive score.

AuthorsOpenReview ↗

Thank You for Your Feedback and Continued Support

Thank you for confirming that our response has addressed your main concerns. We appreciate your constructive feedback, which helped us strengthen the repeatability evaluation, and your continued support.

Reviewer_9ERSOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how risk persists in evolving agent memory. It defines memory residue as source-related risk that survives in generated memory after the original source is hidden.

Critical strengths

(1) The paper studies how risk can persist in agent memory while current outcomes appear safe.

The source-hidden paired design separates the state of memory from the current outcome within model and case, which makes this phenomenon observable.

(2) The paper reports a range of validation checks and releases its materials.

Human audits, lineage swaps, and carrier-dose tests are included, and the cases, prompts, scores, and code are available for inspection.

Critical weaknesses

(1) The paper's central claim of attributing risk across memory updates is not verified by its own experiments.

A. The paired kept/removed contrast only varies the availability of the whole generated episode. Section 3.2 states that it does not isolate the risky semantic component, so the attribution is delegated to an LLM judge rather than established by the design itself.

B. That judge is unstable on the central measurements. Human-coded Execution Propagation is 22.2 pp against 5.6 pp automatically in Section 5.2, and majority aggregation in the Execution rotation panel gives zero Propagation in Section 5.1.

C. The behavioral evidence is close to null. The Task 3 unsafe gaps in Table 2 are +1.1 pp and +0.1 pp, and the Execution Task 2 gap has an interval spanning zero. The central phenomenon is therefore established at the level of labels, not behavior.

(2) The evaluation varies the models but never meaningfully varies the memory, so the findings may not transfer to the memory systems agents actually use.

A. All 18 models share a single minimal memory mechanism, a flat store with a deterministic TF-IDF top-3 retriever and whole-episode writeback (Sections 3.1 and 4.1). The study does not test whether residue forms under the memory systems people deploy, such as consolidating or self-editing designs. The Section 7 adaptations of A-MEM and A-MemGuard cover only six models in small-scale, non-native configurations.

B. The tested memory stores contain only three to five entries with no distractor history (Section 3.1), far below what a running agent accumulates. At this scale top-3 retrieval degenerates and the exposure readout cannot reflect realistic retrieval conditions (Sections 3.2 and 4.3).

Questions for authors
  1. Related to Weakness 1. The main results rest on automatic labels, but Section 5.2 reports human-coded Propagation of 22.2 pp versus 5.6 pp from the automatic labels in Execution. Which of these two numbers represents the paper's actual finding? If the human-coded estimate is correct, do the main conclusions still hold?

  2. Related to Weakness 1. The abstract states that a safe current outcome can coexist with memory that retains source-related risk, but Table 2 shows Task 3 unsafe gaps of +1.1 pp and +0.1 pp. If residue carried into a later task does not measurably change later outcomes, what concrete harm does silent residue pose beyond the memory state itself?

  3. Related to Weakness 2. Appendix F states that hiding source notes does not roll back source-induced context/tag updates in surviving support notes, which can remain in later prompts, and that both arms start from this same source-evolved background. If source-related risk persists in that shared background, what does the kept/removed contrast measure in this adaptation?

3 comments
AuthorsOpenReview ↗

Response to 9ERS (1/2): Attribution, Measurement Reliability, and Behavioral Significance

Thank you for asking us to distinguish attribution, measurement validity, and behavioral significance.

1A — The intervention identifies the effect of G1-carrier availability on source-related residue.

K/R changes access to the whole generated episode on shared history. Strict residue requires source derivation D>=1 and risk preservation R=2 within the same artifact. The abstract, method, conclusion and MC3 carrier-dose description now consistently use this intervention unit. Appendix S checks source/risk judgments against human references on 3,240 semantic artifacts from 1,080 model–case units. These checks validate the source-related risk identified in the resulting memories, complementing the whole-carrier intervention.

1B and Q1 — The two MC1 numbers estimate the same contrast on the same 18 Execution pairs.

The original automatic kept/removed strict-positive counts are 4/18 and 3/18, giving 5.6 pp; the human-reference counts are 6/18 and 2/18, giving 22.2 pp. Two kept false negatives and one removed false positive relative to human coding account for the entire 16.7 pp discrepancy. Thus 91.7% G2 label agreement can coexist with a substantial difference in the paired effect. Both point estimates are positive on this subset; the count reconstruction distinguishes label agreement from effect-size calibration.

Appendix U holds the same 108 historical artifacts and original human reference fixed while reassessing opaque candidates in two reversed presentations. The Execution effects are 16.7 and 22.2 pp, with mean 19.4 [0.0, 41.7] pp and a −2.8 [−8.3, 0.0] pp difference from human coding. The same-material reassessment brings the point estimate closer to the fixed human reference.

The separate 1,080-unit expanded-store panel evaluates calibration at a larger scale:

TrackJ1 Propagation (pp)Human-majority Propagation (pp)J1 − human, 95% CI (pp)
Answer11.8612.05-0.19 [-1.16, 0.79]
Execution4.815.00-0.19 [-0.84, 0.43]

We also report J1/J2/J3 individually on common valid pairs: Answer effects are 11.86/12.04/11.49 pp and Execution effects 4.81/5.19/5.00 pp. Their between-evaluator standard deviations are 0.28 and 0.19 pp. Appendix S provides intervals, label correspondence, joint semantic/behavior disagreements, and the new panel's majority results.

The original Rotation zero is reproduced and explained by the archived votes.

We form a strict label within each evaluator, take the majority separately for each carrier, and then compute the paired contrast. Among the 12 Execution pairs, majority labels give two kept-only positives, two removed-only positives, two positives in both arms, and six negatives in both. Consequently, (2−2)/12=0. Each individual evaluator gives +1/12, but majority voting per carrier need not preserve the sign of those aggregate contrasts. Appendix V reports the vote patterns and the majority interval [−33.3, 33.3] pp. This explains the original aggregation sensitivity; the larger-panel checks are reported separately.

Section 5.2 now distinguishes the historical MC1 calibration issue, its same-material reassessment, and the new panel. Appendices S/U/V and the case-level exports make each calculation traceable.

1C and Q2 — Safe current behavior can coexist with retained source-related risk.

The original Task-3 unsafe gaps are +1.1/+0.1 pp, and the Execution Task-2 interval includes zero. These characterize behavior over the tested horizon. The memory-state result captures a different property: source-related risk remains in persistent content that is eligible for reuse even when a current action is safe. An action guard can suppress activation without removing that retained content.

In the existing case-level analysis, kept-only Propagation is associated with 5.40×/3.27× the risk of kept-only unsafe behavior in Answer/Execution (§4.3). This descriptive association links the memory-state readout to observed behavior and motivates auditing writeback alongside actions. The two readouts reveal what the agent did and what it carries into later tasks; concrete long-horizon harm remains a follow-up question.

Taken together, the expanded human audit and cross-evaluator checks support positive Propagation in both tracks, with close human–automatic effect estimates. The same-material reassessment brings the original Execution estimate closer to the human reference, while the vote-level reconstruction explains the Rotation zero. Together with the observed association between propagation and unsafe behavior, these findings support the reliability and relevance of measuring source-related memory residue alongside current actions. The evidence supports our central claim: safe current behavior alone does not establish that source-related risk has been removed from memory. Supporting results and case-level records are provided in Appendices S/U/V and revision_addendum/E04_measurement_reliability/.

AuthorsOpenReview ↗

Response to 9ERS (2/2): Memory-System Coverage, Selective Retrieval, and Background Sensitivity

Thank you for asking about memory-mechanism coverage, retrieval competition, and source-induced context/tag updates.

2A — The extension covers all 18 models and adds content-changing memory operations.

Each system uses 191 cases per model (108 Answer, 83 Execution), 100 added background memories per case and top-3 actor retrieval. A-MEM covers metadata/link evolution, A-MemGuard retrieval screening, and the Mem0 contract ADD/UPDATE/DELETE/NOOP operations.

SystemAnswer Propagation (pp), 95% CIExecution Propagation (pp), 95% CI
A-MEM10.22 [8.56, 11.84]4.82 [2.87, 6.83]
A-MemGuard3.72 [2.09, 5.49]2.54 [0.94, 4.22]
Mem0 content-update contract9.77 [7.39, 12.23]5.59 [3.43, 7.78]

All six intervals are above zero. Mem0 Silent Residue is 14.31%/10.09%; G2 K/R rates are 40.85%/31.08% (Answer) and 28.40%/22.81% (Execution). UPDATE activity occurs in 68.00%/70.21% of transactions; strict residue in UPDATE records occurs in 25.88%/18.26% of G2-K sets with evaluable UPDATE-specific status. These UPDATE-specific scores extend the residue analysis to records marked as updated.

Mem0 compares executing versus skipping the whole G1 write transaction and uses numerical relevance ranking. A-MEM/A-MemGuard compare G1-carrier availability. The expanded A-MEM analysis focuses on context/tag/link evolution and paired G1-carrier availability; Section 7 separately reports the original six-model embedding-based setup. Appendices R/W specify these controlled contracts, retained mechanisms, changes and their reasons, alongside record IDs, version identifiers and per-record scores. These results extend the positive residue contrasts to metadata evolution, retrieval screening, and content-updating operations.

2B — Selective retrieval retains positive average propagation.

The 18-model reference also uses 100 added memories per case. The expanded reference uses the original TF-IDF function: each retrieval call fits the query together with the currently visible memories and returns the top three. Initial pools contain 103–105 items; Task-2 kept pools contain 103 Answer or 102–103 Execution items after source hiding.

MetricAnswer originalAnswer expandedExecution originalExecution expanded
Formation (%)65.1260.9632.6226.31
Silent (%)21.3519.7011.5910.51
Propagation (pp)12.199.445.763.28
Activation (pp)14.5610.152.811.74
TrackOriginal retrievalExpanded Task-2 G1 Hit@3
AnswerTop-3 returns all available items74.59%
ExecutionTop-3 returns all available items52.88%

Hit@3 measures exposure, whereas Propagation averages all valid K/R pairs, including cases without G1 retrieval. Propagation remains 9.44 [7.63, 11.23] / 3.28 [1.33, 5.22] pp. These two readouts show reduced retrieval exposure together with persistent residue contrasts (Section 4.3; Appendix Q).

Q3 — The rollback checks sensitivity to source-induced background changes.

A separate, internally matched six-model A-MEM panel uses 191 cases per model. The two backgrounds share each task, initial pool and G1 identity; within each background K retains G1 and R removes it. Rollback jointly restores designated source-induced context/tag and associated link updates in surviving notes.

TrackNormal K−RRollback K−RNormal − rollbackValid pairs
Answer13.91 [10.33, 17.61]13.59 [9.73, 17.46]0.32 [-1.54, 2.19]646/648
Execution5.23 [1.82, 8.68]6.05 [2.41, 9.84]-0.81 [-2.63, 1.00]496/498

Values are pp with 95% paired case-cluster intervals. The positive K/R contrast remains after rollback; the between-background changes are +0.32/−0.81 pp. This directly checks the specified shared-background sensitivity. How context, tags and links themselves transmit risk across longer histories remains a further mechanism question.

Section 7 summarizes the system and rollback checks; Appendix T gives its 2×2 design and all four cells. The companion folders revision_addendum/E02_expanded_memory/, E03_memory_systems/ and E05_context_tag_rollback/ provide pool memberships, scores, retrieval records, version changes and recomputation code.

Taken together, positive Propagation persists when retrieval must select among competing memories and across all three tested memory systems, with all six system-level confidence intervals above zero. Rolling back the designated context/tag/link updates also preserves positive contrasts in both tracks. These results show that the memory-residue finding extends beyond near-exhaustive retrieval and a single flat-store mechanism, and remains observable after restoring the specified shared-background updates.

AuthorsOpenReview ↗

Follow-up on Our Responses

Dear Reviewer 9ERS,

Thank you again for your time and thoughtful feedback.

We would like to briefly follow up on our responses, which include expanded-memory experiments, human calibration and evaluator checks, broader memory-system coverage, and the context/tag rollback analysis. We also clarified the original MC1 and Rotation results and the relationship between memory residue and current behavioral outcomes.

With the discussion period closing soon, we would appreciate your feedback on whether these results and clarifications address your concerns. If any questions remain, we would be happy to address them before the discussion closes.

Best regards,

The Authors

Reviewer_wA2mOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper is about measuring memory state separately from immediate behavior. The paired kept/removed design gives a clear diagnostic contrast, and the authors are unusually explicit about several limits of their estimates. My main concern is that the evidence supports a narrower claim than the paper’s framing suggests: it shows effects of making an entire generated episode available in a small, controlled memory store, while the identification of risk-specific propagation depends on a semantic evaluator with known calibration and presentation sensitivities.

Critical strengths

It is valuable for separating stored residue, retrieval exposure, and unsafe outcomes, and for avoiding treating a safe current response as proof that the agent’s memory is safe.

Human audits, lineage and risk-stratum checks, dose tests, evaluator rotations, and repeated transitions probe different parts of the measurement claim. The paper also reports unfavorable or ambiguous results, including weaker Execution effects and substantial A-MemGuard masking failures.

Critical weaknesses

Keeping Gen1 changes what the agent can retrieve, but the carrier includes its task text and other potentially useful content. The resulting contrast combines presentation, response, and writeback. It therefore supports a claim about whole-carrier availability more directly than a claim that risky semantics themselves caused the increase in Gen2 residue. The paper acknowledges this limitation; it should make it central to the framing and avoid implying that the experiment isolates a risk-content mechanism.

The main evaluator sees kept/removed candidate IDs, and kept candidates are presented first. The paper reports an opaque-ID/order replay, but it does not remove this concern for the full census. Since the same evaluator model is used for semantic and behavior judgments, separate prompts do not ensure independent errors. Because the headline propagation estimate is a difference in semantic labels, the authors should report the size and uncertainty of the blinded, order-balanced estimate alongside the primary estimate, and clarify how much the conclusion depends on the exposed-label setup.

Task 2 has at most three available memories and uses top-3 retrieval, so every available item is returned. There are no distractor memories. Thus, the main contrast does not test selective retrieval in a crowded store; it tests whether a generated carrier is present at all. This is useful for a focused diagnostic, but limits how strongly the paper can generalize to evolving memory systems. The paper should foreground this boundary in the abstract and contribution claims.

The strict-positive rule combines source derivation of at least 1 with maximum risk preservation. The human audit supports correspondence, but it is small: 18 actor–case clusters per track. More importantly, human-coded Execution propagation is reported as 22.2 pp versus 5.6 pp automatically, and the paper explains how a few label disagreements shift that paired estimate. This is not merely uncertainty about prevalence; it affects the magnitude of the central effect. The claims should distinguish more clearly between evidence for a positive direction and evidence for the reported effect size.

Questions for authors

Do you think that the Task 3 results are easy to overread?

Do you think the intervention and adapted-system evidence is a bit preliminary?

3 comments
AuthorsOpenReview ↗

Response to wA2m (1/2): Attribution, Evaluation Robustness, and Human Calibration

Thank you for highlighting attribution, presentation sensitivity, and human calibration.

W1 — The paired contrast identifies whole-carrier availability.

On shared history, K/R changes access to the generated G1 episode, including task text and other useful content. Source derivation and risk preservation define the outcome: strict residue requires D>=1 and R=2 in the same artifact. We now use this whole-carrier interpretation throughout the abstract, method and conclusion, and describe MC3 as a carrier-dose manipulation. The human checks validate the source-related risk identified in the resulting memories, complementing the whole-carrier intervention.

W2 — Masked, order-balanced evaluation is supplemented by position reversal and human-referenced error checks.

The expanded-store evaluation uses opaque candidate IDs and counterbalanced kept/removed positions. We then reverse the positions of the same 6,876 G2 items and rescore them with the same primary evaluator and rubric. Both passes mask condition IDs; the second reverses the first pass's positions. The table reports Propagation and its change in pp, with 95% CIs, on common valid pairs:

TrackMasked, balancedPositions reversedChangeValid pairs
Answer9.44 [7.67, 11.25]9.91 [8.10, 11.78]0.46 [-0.26, 1.24]1940/1944
Execution3.28 [1.34, 5.22]3.15 [1.07, 5.22]-0.13 [-0.87, 0.61]1493/1494

To examine shared semantic/behavior errors, we compare both judgments on the same Task-2 arm against H1–H3 majority references. The 1,080 model–case panel includes J1, J2, and J3:

TrackJudgePropagation (pp)Both agree with humans
AnswerJ111.8698.52%
AnswerJ212.0498.61%
AnswerJ311.4998.52%
ExecutionJ14.8199.07%
ExecutionJ25.1998.61%
ExecutionJ35.0097.96%

The last column reports the proportion of items where both semantic and behavior judgments agree with their respective human references. Propagation uses 539/540 common valid model–case pairs in Answer/Execution; joint agreement uses 1,079/1,080 model–case–arm items. Across J1/J2/J3, Propagation SDs are 0.28/0.19 pp. Appendix S reports intervals, the evaluator-comparison figure, and all four categories: both agree, semantic-only disagreement, behavior-only disagreement, and both disagree.

W4 — Three directional label disagreements explain the original effect-size discrepancy.

On the same 18 Execution pairs, the original automatic labels give 4/18 strict-positive kept items and 3/18 removed items, yielding 5.6 pp. The human reference gives 6/18 and 2/18, yielding 22.2 pp. Two kept false negatives and one removed false positive relative to that reference therefore account for the 16.7 pp difference, despite 33/36 G2 labels agreeing.

We retain both original estimates and assess the same 108 historical artifacts with the original human reference fixed. Opaque, position-swapped passes give Execution effects of 16.7 and 22.2 pp; their mean is 19.4 [0.0, 41.7] pp, with a −2.8 [−8.3, 0.0] pp gap to the human reference (Appendix U). The same-material reassessment brings the point estimate closer to the fixed human reference; the larger, separately sampled panel then tests calibration across all 18 models:

TrackJ1 Propagation (pp)Human-majority Propagation (pp)J1 − human, 95% CI (pp)
Answer11.8612.05-0.19 [-1.16, 0.79]
Execution4.815.00-0.19 [-0.84, 0.43]

The revised presentation distinguishes the original discrepancy, same-material reassessment, and larger-panel calibration. Together, these checks address agreement on the paired effect rather than relying on label agreement alone. Section 5.2 summarizes the findings; Appendix S reports the new panel and Appendix U the original-panel diagnostic. Case-level scores and the corresponding aggregation code accompany the tables.

The reproducible tables are in revision_addendum/E04_measurement_reliability/.

Taken together, Propagation remains positive under masked, order-balanced evaluation and position reversal, with closely aligned estimates across three evaluators and the human reference. Joint semantic/behavior disagreements are rare on the audited panel, and the same-material reassessment brings the original Execution estimate closer to human coding. These results support the reliability of the source-related residue measurement under the whole-carrier availability contrast.

AuthorsOpenReview ↗

Response to wA2m (2/2): Selective Retrieval, Behavioral Interpretation, and Memory-System Coverage

Thank you for asking about selective retrieval and the interpretation of the Task-3, intervention and memory-system results.

W3 — Positive propagation persists when top-3 must select among competing memories.

We add 100 non-source background memories per case across all 18 models and 191 cases, retaining the top-3 budget. Initial pools contain 103–105 entries; after source hiding, Task-2 kept pools contain 103 Answer or 102–103 Execution entries. The expanded reference uses the original TF-IDF function: each retrieval call fits the query together with the currently visible memories and returns the top three.

MetricAnswer originalAnswer expandedExecution originalExecution expanded
Formation (%)65.1260.9632.6226.31
Silent (%)21.3519.7011.5910.51
Propagation (pp)12.199.445.763.28
Activation (pp)14.5610.152.811.74

Task-2 G1 Hit@3 directly measures retrieval exposure: the percentage of kept branches retrieving G1 among the three selected items.

TrackOriginal retrievalExpanded G1 Hit@3
AnswerTop-3 returns all available items74.59%
ExecutionTop-3 returns all available items52.88%

Across all valid K/R pairs, including cases without G1 retrieval, Propagation remains 9.44 [7.63, 11.23] / 3.28 [1.33, 5.22] pp (Answer/Execution; 95% case-cluster intervals). Reduced exposure and positive residue contrasts jointly address the near-exhaustive-retrieval concern. Section 4.3 summarizes the findings; Appendix Q provides configuration comparisons, model-level results and retrieval definitions.

Q1 — Task 3 measures the tested behavioral consequence of G2 availability.

Task 3 retains the same G1 in both arms and varies G2-K availability. The original unsafe gaps, +1.1/+0.1 pp, are small effects over that horizon. The memory-state finding is that source-related risk can remain available while current actions are safe. Action checks measure what the agent did; writeback checks measure what remains available for later reuse. Section 4.3 states these complementary roles. Concrete longer-horizon harm remains a follow-up question.

Q2 — Interventions are diagnostic; the memory-system extension now covers all 18 models.

Section 6, “Intervention Diagnostics,” retains its original six-model configuration. Safety-prompt and gate comparisons characterize partial control within the tested protocol, relating intervention, residue and current behavior. Their role is to diagnose this relationship, rather than establish a complete defense.

For system coverage, each integration uses all 18 models and 191 cases, 100 added memories per case and top-3 actor retrieval. A-MEM represents metadata/link evolution, A-MemGuard retrieval screening, and Mem0 content-changing ADD/UPDATE/DELETE/NOOP transactions:

SystemAnswer Propagation (pp), 95% CIExecution Propagation (pp), 95% CI
A-MEM10.22 [8.56, 11.84]4.82 [2.87, 6.83]
A-MemGuard3.72 [2.09, 5.49]2.54 [0.94, 4.22]
Mem0 content-update contract9.77 [7.39, 12.23]5.59 [3.43, 7.78]

All six residue contrasts are positive with intervals above zero. Mem0 Silent Residue is 14.31%/10.09%; absolute prevalence varies across systems. Mem0 compares executing versus skipping the whole G1 write transaction; A-MEM/A-MemGuard compare carrier availability. Appendices R/W specify the retrieval implementations, retained operations and integration changes for these controlled comparisons.

Section 7 summarizes the additions. Appendices R/W give complete system results, absolute rates and operation diagnostics. The companion folders revision_addendum/E02_expanded_memory/ and revision_addendum/E03_memory_systems/ provide pool memberships, model–case scores, record-level data and recomputation code.

Taken together, the expanded-store results show positive Propagation despite substantially reduced G1 retrieval, and all three tested memory systems retain positive contrasts in both tracks, with confidence intervals above zero. The residue finding therefore extends beyond near-exhaustive retrieval and a single memory mechanism. Task-3 and intervention results complement this finding by showing why current behavior and retained memory should be assessed together.

AuthorsOpenReview ↗

Follow-up on Our Responses

Dear Reviewer wA2m,

Thank you again for your time and constructive feedback.

We would like to briefly follow up on our responses, which include masked and order-balanced evaluation, cross-evaluator and human-reference checks, expanded-memory experiments, and broader memory-system coverage. We also clarified the whole-carrier interpretation and the roles of the Task-3 and intervention results.

With the discussion period closing soon, we would appreciate your feedback on whether these results and clarifications address your concerns. If any points remain unresolved, we would be happy to provide further clarification before the discussion closes.

Best regards,

The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions and Responses to Reviewer Concerns

We thank the reviewers and Area Chair for their time and constructive feedback.

The revised manuscript and additional evaluations strengthen the central finding: source-related risk can persist in memory despite safe current behavior, and the measured propagation persists across independently regenerated histories, competitive retrieval, and multiple memory mechanisms. We summarize the main evidence below; the reviewer-specific responses provide the detailed explanations and results.

1. The finding persists across full-history regeneration. Using the original initial memory pools, we repeated the complete Task-1-to-Task-3 protocol five times across all 18 models and 191 cases, freshly generating G1 and G2 in each run. Mean Propagation was positive in both tracks in every run: 11.94 ± 0.69 pp for Answer and 6.11 ± 1.23 pp for Execution (mean ± run-level SD). These results address dependence on a fixed set of generated parent memories. Reviewer 5W79 has confirmed that our response addressed their main concerns. See §5.1 and Appendix P.

2. Positive propagation remains when top-3 retrieval must select among competing memories. With 100 additional background memories per case and the original retrieval method, Task-2 G1 Hit@3 falls to 74.59%/52.88% in Answer/Execution. Propagation remains 9.44/3.28 pp, with both 95% confidence intervals above zero. The propagation estimates include valid pairs where G1 was not retrieved. Thus reduced exposure coexists with a persistent memory-residue contrast, addressing the near-exhaustive-retrieval concern. See §4.3 and Appendix Q.

3. The results extend across memory mechanisms on the full 18-model panel. A-MEM, A-MemGuard, and the Mem0 content-update contract cover metadata/link evolution, retrieval screening, and ADD/UPDATE/DELETE/NOOP operations, respectively. Their Answer/Execution propagation contrasts are:

Memory systemAnswer (pp)Execution (pp)
A-MEM10.224.82
A-MemGuard3.722.54
Mem0 content-update contract9.775.59

All six 95% confidence intervals are above zero. Each system uses the expanded memory pools. The corresponding interventions are defined for each mechanism: carrier availability for A-MEM/A-MemGuard and whole-G1-write-transaction retention for Mem0. See §7 and Appendices R/W.

4. Human calibration and evaluation controls support the measurement. On the separate 1,080 model–case audit panel, automatic and human-majority Propagation estimates are 11.86 versus 12.05 pp for Answer and 4.81 versus 5.00 pp for Execution. J1/J2/J3 estimates are closely aligned, and joint semantic/behavior disagreements with human references are rare on the audited panel. Propagation also remains positive under masked, order-balanced evaluation and position reversal. We trace the original MC1 discrepancy to three directional label disagreements, reassess the same historical materials against the fixed human reference, and reconstruct the original Rotation zero from its per-item votes. These analyses separately address calibration, presentation sensitivity, shared errors, and aggregation. See §5.2 and Appendices S/U/V.

5. Targeted controls clarify attribution and background sensitivity. The core paired intervention changes whole-carrier availability; source derivation and risk preservation define the measured residue. MC3 is correspondingly described as a carrier-dose manipulation. On the matched six-model A-MEM panel, positive propagation remains after jointly restoring the designated source-induced context/tag/link updates. See §7 and Appendix T.

6. Behavioral and intervention results explain the value of the memory-state readout. Task 3 measures behavior over the tested horizon, while writeback assessment identifies retained content available for later reuse. In the existing case-level analysis, kept-only Propagation is associated with 5.40×/3.27× the risk of kept-only unsafe behavior in Answer/Execution (§4.3). This descriptive association supports examining memory alongside actions. Section 6 frames SafePrompt and gate interventions as diagnostics of how residue, exposure, and current behavior respond within the tested protocol.

Together, these additions support the repeatability, measurement reliability, and cross-configuration scope of the memory-residue finding. They reinforce the value of assessing retained memory alongside current actions. The updated PDF highlights revisions in blue; Appendices P–W provide the detailed analyses, and the supplementary archive's revision_addendum/ modules E01–E05 provide the corresponding data and analysis code.

With the discussion period closing soon, we welcome any remaining questions and would be happy to provide further clarification.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.