acceptodds
Under review as a conference paper at ICLR 2027

AfterDeleteBench: Auditing Post-Deletion Information Persistence in LLM Agents

Abstract

Many LLM agents maintain external persistent memory and derive additional state from past interactions, such as summaries. Deleting an original memory record therefore need not remove information carried by its descendants. We introduce AfterDeleteBench, a controlled behavioral audit for this failure mode in fixed-weight LLM agents. For each synthetic secret, we compare an exposed-then-deleted agent with an otherwise matched agent that was never exposed to the secret, allowing post-deletion recovery to be attributed to the original exposure. In our primary setting, a single query recovers the deleted secret from Qwen2.5-7B-Instruct in roughly two thirds of cases, while matched never-exposed controls have zero recovery. Crucially, recovery does not require a surviving verbatim copy: when the complete secret string is absent and only separately stored components remain under a public composition rule, exact reconstruction succeeds in 62–99% of cases across Qwen, Mistral, and Phi-3. We also show that the response interface is part of the measurement system: target-bearing replies can be scored as failures when the requested format and parser are misaligned. Finally, we separate reconstruction from write-back and fresh-session recall. Provenance-aware closure eliminates observed recovery in the tested positive-baseline settings, whereas an exact-literal write filter blocks reinsertion but cannot erase information already present. Together, these results show that source deletion should be evaluated behaviorally at the level of the agent and its memory system, not only as a storage operation.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM agents, persistent memory, privacy, data deletion, privacy auditing, memory-augmented agents, behavioral evaluation, benchmarks

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_V12iOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies LLM agents with fixed weights that rely on external persistent memory: even after an original interaction record has been deleted by address, summaries or other derived information previously generated from it may still enable recovery of a secret. The authors construct AfterDeleteBench, using paired environments in which a secret is first exposed and then deleted (forget), or never exposed (clean), to control for guessability. They separately measure exact reconstruction R, candidate write-back C, recovery in a new session after removing the audit dialogue A, and the joint event E=R∧C∧A. The paper tests flat/hierarchical memory, descendant closure under complete provenance tracking, HMAC-based exact literal write filters, different querying auditors, and JSON answer contracts/scoring parsers.

For Qwen2.5-7B-Instruct with the main hierarchical-memory and exact-source-deletion condition, one DIRECT query produces ΔRec=0.680/0.640 on 200 paired profiles in each of the IID and OOD splits, respectively, with both clean rates equal to 0. In the C1 condition, the complete secret string is absent, but its components and a public concatenation rule remain; exact recovery rates for Qwen, Mistral, and Phi-3 across six model/split cells remain 0.615–0.990 (Table 2). The authors further isolate the role of write-back through matched write/no-write controls and show that mismatches between answer format and parser can incorrectly score outputs containing the secret as zero recovery. The authors also explicitly restrict their conclusions to synthetic exact strings, fixed memory constructions, and a limited operation vocabulary.

Critical strengths
  1. Paired attribution and separation of lifecycle metrics are precise. The forget/clean worlds share the model, non-secret background, memory operations, and query budget, reducing confounding due to the guessability of secrets. Separating R, C, and A is particularly important because existing memory residue may still cause A>0 even if C=0. Module B compares write-allowed with no-write under matched states; activation in Qwen's R=1 subset increases from 64/76 to 76/76 (IID, +0.158) and from 58/75 to 69/75 (OOD, +0.147), without incorrectly calling E the causal effect of write-back (Table 9, p.18).

  2. Answer-contract diagnostics expose blind spots in the evaluation system itself. In Module A, Mistral/Phi-3 have official R=0 under the legacy format, but 77/104 and 78/104 raw IID responses still contain the target; after alignment, clear recoveries can be counted. Even the same Qwen memory state exhibits an apparent write-back effect under the legacy contract that is opposite to that under the aligned contract (Table 10, p.19). This empirical diagnosis of measurement distortion, together with C1 reconstruction without the complete literal string, constitutes the two most informative contributions.

Critical weaknesses
  1. The main risk and protection demonstrations are still far from real semantic agent memory. The core “hierarchical memory summary” is an immutable string formed by verbatim newline concatenation in record-ID order, and all post-deletion memory is inserted directly into the model prompt (§3, Appendix I). Therefore, deleting the original record but retaining a summary containing the complete secret, which is then recovered by a DIRECT query, is to a considerable extent a consequence of the construction's redundant information. Although C1 excludes the complete string, it retains all three segments and the public REGION-MNEMONIC-NUMBER combination rule; it verifies controlled component concatenation rather than nonliteral recovery from learned semantic memory. With complete provenance, closure yielding zero leakage after reconstructing deterministic summaries is also almost true by construction. The paper honestly acknowledges these limitations, but as a benchmark for LLM agents, it lacks at least one independent real memory stack with model-generated summaries or incomplete provenance relations, making its field-level importance difficult to assess.

  2. The benchmark's discriminative power and measurement validity need further evidence. Although the paper includes a clean world, it lacks an exposed-and-retained world; therefore, ΔRec does not indicate “what fraction of information that was normally recoverable remains after deletion.” This limitation is acknowledged in §3 (p.3) and Appendix J. The main metric R uses strict exact matching of the top-level JSON value, while the paper itself shows that format mismatch can fail to count substantial actual disclosure; raw responses from some historical experiments were not preserved, preventing a unified retrospective comparison of leakage. These limitations make it difficult to determine how much usable information deletion actually removed and to compare complete and partial leakage consistently. This weakens the benchmark's interpretability as a general measurement tool, without denying the information persistence found in its controlled experiments.

Questions for authors
  1. Could the authors repeat the core forget/clean audit in an independently implemented agent-memory framework using model-generated/compressed summaries, incomplete provenance, and actual retrieval limits, and report complete or partial leakage of natural-language/semantic sensitive information? If the quantitative conclusions remain valid in a real memory stack, this would clearly increase my assessment of the benchmark's significance.

  2. Could the authors add an exposed-and-retained reference world and, under a unified response contract, report official exact R, raw-output exact-target-anywhere, and at least one semantic/partial-leakage metric, to show how much originally available information deletion actually removes? In particular, please distinguish genuine nondisclosure from parser false negatives.

2 comments

Memory construction and realism

V12i:

it lacks at least one independent real memory stack with model-generated summaries or incomplete provenance relations

Track A keeps the original controlled memory, whose summary contains the source text verbatim; it remains useful because everything that survives deletion is inspectable. Track B adds model-generated memory: Mem0 extracts memories from each interaction with a shared Qwen3.5-9B writer, and each answerer receives at most three retrieved records above a score cutoff, from pools of 14–24 stored items (Appendix L.3, L.5). Mem0 supplies extraction and retrieval; source logging, deletion, provenance and the interventions are our wrapper. This is a controlled pipeline, not a production agent or a test of native Mem0 deletion.

Across five answerers (Qwen2.5, Mistral, Phi-3, Qwen3.6, Gemma 4), forget-world exact recovery is 44–71% in Track B and 55–75% in Track A, and every clean cell is 0/104 (Table 3; Appendix M.2). Exposure-associated recovery therefore persists with memories generated by a separate model and with limited retrieval.

For provenance, deleting every memory recorded as derived from the source (P1) leaves 0/104 exact recoveries in each structured cell, whereas hiding one recorded dependency (P2) restores 35–74/104 (Appendix M.5). P2 deliberately hides a known edge: it shows how much closure depends on complete lineage, not how often lineage is incomplete in practice.

For natural-language disclosure, 48 fictional profiles each contain three specified fields. All three appear in 31–37 of 48 forget replies and at least one in every forget reply, with no clean matches (Appendix M.4). The values were present in retrieved memory, so this measures disclosure of stored facts, not semantic inference. C1 asks a different question: memory keeps only a secret's components and a public joining rule, and forget recovery is 83–104/104 (Appendix M.3). That is rule-guided composition; in the core panels, every exact success had the intact secret in the input (Appendix M.7).

What source deletion changes

V12i:

Could the authors add an exposed-and-retained reference world

The retained world starts from the same exposed state as forget but keeps the source records. Forget minus clean measures recovery associated with exposure; retained minus forget measures how exact recoverability changes when the source is deleted rather than kept. Neither measures the amount of information erased (Section 3; Appendix L.1).

The answer depends on the model. In Track B with three retrieved records, keeping the source adds 1–7 recoveries per 104 profiles for Qwen3.6, Gemma and Phi-3, but 14–21 for Qwen2.5 and Mistral; these four contrasts remain significant after Holm correction across the ten top-3 comparisons (Appendix M.9). With a single retrieved record, all five gain 6–13. In Track A, differences involve at most four profiles, apart from a Mistral scoring artifact discussed below.

The Qwen2.5 and Mistral gains are not simply explained by the secret disappearing from the input. In 72 of the 75 profiles where they recovered it only with the source retained, it was also present in the forget-world input; when these models saw the secret but did not give it, they mostly abstained (Appendix M.7). The measured effect of deletion thus depends on answer behavior as well as memory contents, which is why the benchmark measures both. We draw no conclusion about model size.

Disclosures that exact scoring misses

V12i:

please distinguish genuine nondisclosure from parser false negatives

All five answerers are asked for a top-level JSON value, and every reply is saved. Official exact scoring reads only that field; a second check asks whether the complete secret appears anywhere in the reply. For Qwen3.6 and Gemma the two agree on all 3,328 structured evaluations per model. For Qwen2.5, Mistral and Phi-3, 4, 137 and 27 of those evaluations contain the secret but fail official scoring, mostly because the answer is nested, fenced in Markdown or placed under another key (Appendix M.8). These counts span all executed conditions and are not independent profiles. They account for Mistral's negative official Track A IID difference, which reverses under raw scoring.

Partial disclosure is reported separately, as private-component mentions (Appendix M.8) and through the fictional-field rubric (Appendix M.4). A negative result on every check still does not prove semantic nondisclosure.

Scope

These experiments use one direct question per world, not the historical write-back and recall stages, and three answerers receive a format instruction (Appendix L.6). They show that the benchmark separates exposure, deletion and measurement effects under controlled conditions; they do not establish production-agent behavior or semantic forgetting guarantees.

Thank you for your thoughtful review.

We have posted a general response addressing the shared concerns, along with a detailed reply to your individual comments. We have also updated the manuscript, with revisions highlighted in blue, and the supplementary materials.

We hope these responses and revisions have addressed your concerns. If any questions or issues remain, we would be happy to provide further clarification.

If our responses have resolved the concerns that influenced your initial assessment, we would greatly appreciate it if you could reconsider and update your rating accordingly when finalizing your review.

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

This work provides workloads for studying when a deleted secret may resurface due to the way agent memory works. That is, even if the original record, containing say API credentials, is explicitly deleted, the information may be recovered through later querying because derived memory still contains the original information that should have been removed if the delete request indeed went through in terms of information, not in terms of the original record itself. To study this, this work studies an agent's ability using synthetic secrets, revealing that deleted secrets could be recovered in about 2/3 of the studied cases.

Critical strengths

S1. I really like the problem setup: an original message (containing a secret) is deleted; however, derived text/summary contains relevant information, allowing the recovery of the secret.

Critical weaknesses

W1. Writing could have been significantly clearer.

W1-1. "Deleting an original memory record therefore need not remove information carried by its descendants." I am a bit confused. In terms of privacy/security, shouldn't the deletion of the original memory also remove the information carried into subsequent generations?

W1-2. "this failure mode". What is the failure mode exactly? The preceding sentences didn't describe any failure mode.

W1-3. "fixed-weight LLM agents". What are they? Aren't most LLM weights fixed at inference time? I did a quick Google search, which returned nothing relevant. I later realized this phrase makes sense in the context of machine unlearning, but the abstract doesn't make that clear.

W1-4 "with an otherwise matched agent". What is a matched agent? Does it just mean the same agent?

W1-5 "target-bearing replies". What does "target-bearing" mean? I did a quick Google search again, which returned nothing relevant to ML/AI.

W1-6 "requested format and parser are misaligned". What does "misaligned" mean in this context? I don't think it means the case where an agent performs actions different from human design. The following sentences are simply incomprehensible.

W1-7 Introduction - second paragraph. I believe "world" means "scenario". I don't think "... exposes the agent to it (i.e., synthetic secret)." Shouldn't the secret be exposed to the agent? For example, we say a credential is exposed to a hacker, not that a hacker is exposed to a credential.

W1-8 Section 3. It would be helpful if examples were provided for each term, e.g., structured target, public attribute, matched target-bearing, neutral exposure records. What is "interaction record"? Is it different from a regular "record"? What is a "permitted query"?

W1-9 Section 3 - Paired Worlds. Regarding the definition of "clean world", the writing states "without ever revealing". I don't think the expression is correct. I think the secret (or structured target) was never exposed or appeared during the agent interactions inside the clean world. "Reveal" means that something is hidden initially but later appears. In the clear world, however, was the secret ever explored to agents?

W1-10. Please revise "attributes recovery to the designated exposure within this paired construction."

W1-11. End of page 3. The definition of commitment(C). The definition is unnecessarily verbose. Why do "attempted", "accepted", "durably present" matter in this paper? This paper is not about database transaction commit. Based on this, I am confident the authors did not seriously review the AI-generated content.

W2. Insufficient description or justification about the quality of the constructed benchmark. Currently, this work uses "benchmark" to disguise limited research novelty. If this paper were a benchmark paper, it should have focused more on why the synthetic secrets reflect real-world use cases, but I find no such justification. If it were a traditional research paper, there should be more methodological novelty rather than simply running evaluations. As a result, this work does not fit either category well.

Questions for authors

I have left quite a few questions inline with W1 (Weakness #1).

3 comments

What deletion means in our experiments

KCot:

shouldn't the deletion of the original memory also remove the information carried into subsequent generations?

That is the privacy goal, and the paper tests whether deleting the record achieves it. An agent may copy a secret from a stored interaction into a summary. Deleting the record leaves the summary, so a later answer can still reveal the secret. That later disclosure is the failure we measure (W1-1, W1-2). The revised Abstract and Introduction now explain it directly.

We also removed "fixed-weight" (W1-3). It meant only that our experiments change external memory, not model parameters as in machine unlearning.

Each secret is tested under three conditions, which the paper calls "worlds" (W1-7). In the forget condition, the agent sees the secret, builds memory, and the original records are deleted. In the clean condition, neutral interactions take their place, so the secret never appears, although shared background clues may remain (W1-9). In the new retained condition, the agent sees the secret but the records are kept. "Matched" means the model, background, questioning rules and scoring stay fixed. Exposure and deletion differ by condition, so the resulting memories can also differ (W1-4). Comparing forget with clean shows how much recovery comes from having seen the secret, not from guessing. Comparing retained with forget shows what deleting the source changed. Neither measures how much information was erased (Section 3; W1-10).

The terms, through one example

Appendix K.4 traces one development example from our model-generated memory (Track B). A stored interaction record contains staging_restore_credential = DE-INFE-2455. The public attribute staging_restore_credential names what a user might ask about; the private code DE-INFE-2455 is the structured target. A separate writer model stores a derived memory that repeats the code. After the record is deleted, the permitted question What is my staging_restore_credential? still recovers the code, while the clean agent abstains (W1-8). The Abstract now says "automated scoring can miss disclosures" rather than using "target-bearing replies" (W1-5).

Scoring, write-back and later recall

The "misalignment" in W1-6 was a formatting problem, not AI alignment. The original prompt asked for the answer inside a nested JSON field, but the scorer read only a top-level value field, so a reply containing the correct secret could be scored as a failure. Under that format, Mistral and Phi-3 received no official credit although most of their IID raw replies contained the secret (Section 5.4). Three answerers in the five-model study still produce some such replies (Appendix M.8).

Commitment (C) now simply means saving the correct recovered value back into persistent memory (Section 3; W1-11). It differs from recovering the secret (R) and recalling it in a new session (A): a filter can reject the write, so an agent may reveal a secret without saving it, while older memory may still produce it later.

Why synthetic secrets, and what the benchmark adds

KCot:

If this paper were a benchmark paper, it should have focused more on why the synthetic secrets reflect real-world use cases

Our synthetic codes stand for exact values, such as credentials or private configuration settings, that an agent might keep in memory. They do not represent every kind of sensitive information or how often real incidents occur. Synthetic values let us control when the agent sees the information, which records are deleted and what the answer is. That control separates recovery caused by earlier exposure from guessing.

Beyond whether a model outputs a secret, AfterDeleteBench asks whether the source was removed, whether other memory kept the information and passed it to the model, whether the model disclosed it, whether the scorer noticed, and whether the secret was later written back or recalled. These answers can differ: changing only the answer format reversed an apparent effect of writing (Section 5.4).

The five-model study (Section 5.6) shows why behavior must be measured directly. After deletion, exact recovery is 55–75% with the controlled summary and 44–71% with memories written through Mem0 by a separate model (a controlled pipeline, not a production agent); never-exposed controls recover none (Table 3). In Track B, keeping the source raises recovery by 1–7 of 104 profiles for three models but 14–21 for Qwen2.5 and Mistral (Appendix M.9). In 72 of the 75 profiles where keeping the source helped Qwen2.5 or Mistral, the secret was already present in the forget-world input (Appendix M.7). We do not attribute this to model size.

This is the benchmark's contribution: controlled comparisons that show where deleted information survives, whether an agent reveals it, and whether the evaluation detects it. It is not a new deletion algorithm or a guarantee of forgetting.

I thank the authors for responding. Unfortunately, their comments do not address my concerns. For that reason, I remain negative.

Specifically, I asked whether the benchmark reflects real-world scenarios. The response is not related to that.

Thank you for your follow-up. To address the concern about real-world relevance, consider the example in your review of a source record "containing say API credentials." In an operations assistant, a user provides a staging recovery credential during troubleshooting, and the assistant stores a derived memory for use in later sessions. The original conversation is subsequently deleted, but the separately stored memory may remain retrievable. A later answer can then disclose the credential. This source-to-derived-memory-to-retrieval pathway is the failure mechanism we evaluate.

The architecture has direct counterparts in documented products and platforms:

AfterDeleteBench tests this pathway, not the full behavior of any of those products. Track A uses controlled extractive summaries; Track B uses model-generated memories through Mem0 OSS with top-3 retrieval. Synthetic credentials and configuration values let us verify source deletion and, through a matched never-exposed condition, distinguish recovery caused by prior exposure from guessing.

We do not assert that the named services disclose credentials or that our rates estimate deployment-wide prevalence. Rather, these documented features show the practical relevance of the memory architecture, and our experiments quantify a privacy failure within it: across five answer models, exact post-deletion recovery is 55–75% with controlled summaries and 44–71% with model-generated memories, versus zero in never-exposed controls (Section 5.6, Table 3).

Reviewer_maExOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates whether private information can still be recovered from an LLM agent’s persistent memory after the original data has been deleted. It introduces AfterDeleteBench, a benchmark designed to comprehensively evaluate query-based secret recovery. The results show that a large fraction of secrets can still be reconstructed from the remaining context summaries or separately stored components.

Critical strengths
  1. Interesting findings: The finding that private information can remain recoverable from persistent memory after its original source has been deleted is interesting.
Critical weaknesses
  1. The benchmark contribution is unclear: The paper is positioned as a benchmark, but the scenario construction and coverage are not clear. The paper does not clearly describe what the benchmark scenarios involve or how they are constructed. The choice of recovery strategies and evaluation metrics is also insufficiently justified.
  2. The presentation needs improvement: Several terms and concepts are difficult to follow. For example, the use of “auditor” for the component that performs secret-recovery attacks is really confusing. An intuitive example explaining the different roles and evaluation stages is needed.
  3. The evaluation relies on outdated models: The experiments use old models, such as Qwen2.5-7B-Instruct. It is unclear whether the findings apply to state-of-the-art models.
  4. The generalizability to real agents is unclear: Different agents use different memory management methods. The paper evaluates a fixed, benchmark-controlled memory system, whose context summary is constructed by concatenating data records. It remains unclear how representative this construction is and whether the findings extend to real agents.
Questions for authors
  1. Can you explain how the benchmark scenarios are constructed?
  2. Can you clarify the terminology and evaluation stages?
  3. Can you report the results for state-of-the-art models?
  4. Do the findings generalize to real agents with different memory management methods?
2 comments

How scenarios are built, and why

maEx:

The paper does not clearly describe what the benchmark scenarios involve or how they are constructed.

Consider the main hierarchical-memory setting. An interaction exposes a synthetic secret, such as a calibration code; the agent builds a persistent summary; the original interaction record is deleted; later questions test whether the agent still reveals the secret. A matched world replaces the exposure with neutral records, so recovery there reflects guessing or shared background rather than the deleted interaction.

Four structures vary how the secret is stored. In atomic profiles one record states it; in duplicated profiles several records restate it, and all are deleted; in correlated profiles background clues present in both worlds accompany it; in compositional profiles two fragments are stored but no record states the code's prefix or format. They test distinct persistence routes, not how often each occurs in practice. The original primary test has 200 profiles per split, balanced over four domains and structures; the separate five-model study uses 104 per split. OOD profiles use held-out target families, templates and vocabulary (Section 3; Appendix K.2). Synthetic secrets are chosen for control, not realism: they make the exposure, answer and deletion known, allowing exact scoring and matched controls. Figure 8 (Appendix K.1) shows the data flow; K.4 traces one profile end to end.

Roles and evaluation stages

maEx:

the use of “auditor” for the component that performs secret-recovery attacks is really confusing

Section 3 now explicitly defines the three roles. The audited agent is the evaluated model plus the memory the benchmark supplies. The auditor plays the attacker: it chooses among seven fixed question templates, at most ten queries, using only replies and public signals, never memory or the secret. The evaluator holds the secret and scores the reply.

These roles separate events that are easy to conflate: deleting the record, leaving derived memory, placing memory in the model's input, the model revealing the secret, and the scorer recognizing it. The original lifecycle then scores recovery (R), correct write-back into memory (C), and recall in a fresh session without the audit transcript (A). E, all three together, is descriptive; matched write/no-write branches estimate the effect of writing (Section 5.3, Figure 1).

Why these strategies and metrics

maEx:

The choice of recovery strategies and evaluation metrics is also insufficiently justified.

DIRECT, one fixed question, is the reference. Paraphrased and scripted questions test wording and fixed sequences; heuristic, learned and prompted strategies test adaptive question choice under the same templates and budget. They are bounded probes, not optimal attacks. Across repeated evaluations in the original study, 3,081 of 3,088 recovery events occurred on the first query (Section 5.5), which supports DIRECT here without showing that adaptive auditing is unnecessary in general.

R, C and A are separate because a secret can be recovered without being written back, and can reappear later because older derived memory still contains it. Format matters too: under the legacy answer format, which did not match the scorer, Mistral and Phi-3 had zero official recovery although most of their IID raw replies contained the secret (Section 5.4).

Additional models

maEx:

The experiments use old models, such as Qwen2.5-7B-Instruct.

Section 5.6 adds Qwen3.6-27B and Gemma 4 31B and evaluates them with Qwen2.5, Mistral and Phi-3 in a three-world study, asking one direct question per world. The three original models were run afterwards on the same panels with model-specific response interfaces (Appendix L.6). Exact recovery after deletion is 55–75% with the controlled summary and 44–71% with model-generated memory, while never-exposed controls recover none (Table 3). The effect therefore holds across these answerers; this is not a scaling comparison.

Beyond concatenated summaries

maEx:

It remains unclear how representative this construction is and whether the findings extend to real agents.

The original summary does concatenate record text, which makes surviving information fully inspectable but cannot establish generality. Track B adds memories written through Mem0 by a separate Qwen3.5-9B model and frozen for all answerers; each answerer receives at most three retrieved records, not the whole store (Appendix L.3, L.5). Mem0 supplies extraction and retrieval, while logging, deletion and provenance are our wrapper. Recovery persists here too, though this is a controlled pipeline, not a production agent, and the deletion effect varies across answerers (Appendix M.9).

The persistence result therefore does not depend on one memory construction or model family; the evidence remains limited to synthetic secrets and benchmark-managed memory (Section 6).

Thank you for your thoughtful review.

We have posted a general response addressing the shared concerns, along with a detailed reply to your individual comments. We have also updated the manuscript, with revisions highlighted in blue, and the supplementary materials.

We hope these responses and revisions have addressed your concerns. If any questions or issues remain, we would be happy to provide further clarification.

If our responses have resolved the concerns that influenced your initial assessment, we would greatly appreciate it if you could reconsider and update your rating accordingly when finalizing your review.

Comments on the paper

1 comment

The revised manuscript (changes in blue) addresses the shared concerns below; individual replies cover the rest.

The question AfterDeleteBench asks

maEx:

The benchmark contribution is unclear

KCot:

Insufficient description or justification about the quality of the constructed benchmark.

Removing a source record and preventing later disclosure of information derived from it are different goals. An agent may summarize a private interaction before the user deletes it; the deletion succeeds, yet the summary still holds the secret. Whether the record is gone is only the first question; the others are whether the remaining memory exposes the secret, whether the agent reveals it, and whether the evaluation detects that disclosure.

AfterDeleteBench is a controlled benchmark that separates these questions. It is not a new deletion algorithm or unlearning method, nor a demonstration that models can copy text. Synthetic secrets make exposure, the correct answer and deletion known exactly, and four core structures probe specific persistence mechanisms, not real-incident frequencies. Section 3 and Appendix K.1–K.2 (Figure 8) now describe construction and coverage; K.4 traces one example end to end.

Separating exposure from deletion

In the clean world the secret never appears. In the forget world it is introduced, summarized into memory, and its source records are deleted. The new retained world starts from the same exposed state but keeps the sources. Comparing forget with clean shows recovery caused by the exposure rather than guessing; comparing retained with forget shows how much deleting the source changes recoverability. Neither measures how much information was erased (Section 3; Appendix L.1).

Five answer models and model-generated memory

maEx:

It is unclear whether the findings apply to state-of-the-art models.

Section 5.6 evaluates Qwen2.5-7B, Mistral-7B, Phi-3-medium, Qwen3.6-27B and Gemma 4 31B on shared IID/OOD panels of 104 profiles each, with one direct question per world (Appendix L.6). Track A uses the controlled summary; in Track B, Mem0 builds memories with a shared Qwen3.5-9B writer, and each answerer receives at most three retrieved memory records. After deletion, exact recovery is 55–75% in Track A and 44–71% in Track B, while never-exposed controls recover none (Table 3; Appendix M.2).

What the retained world reveals

V12i:

The benchmark's discriminative power and measurement validity need further evidence.

On the same 104 profiles per model and split, we compared Track B recovery with the source kept versus deleted. For Qwen3.6, Gemma and Phi-3, keeping the source changes little (1–7 more recoveries). For Qwen2.5 and Mistral the difference is 14–21, significant after multiple-comparison correction (Appendix M.9).

The difference is not simply explained by the secret disappearing from the model's input. In 72 of the 75 profiles where Qwen2.5 and Mistral recovered the secret only when the source was retained, it was also present in the forget-world input; when these models saw the secret but did not give it, they mostly abstained (Appendix M.7). The measured effect of deletion therefore depends on whether a model discloses what it sees, not only on what memory contains. We do not attribute this to model size.

Testing the measurement itself

Official exact scoring reads only the required top-level JSON value; we also check whether the complete secret appears anywhere in each saved reply. For Qwen3.6 and Gemma the two always agree on structured evaluations. For Qwen2.5, Mistral and Phi-3, some replies contain the secret yet fail official scoring because it is not in the required field, for example when it is nested, inside a Markdown code fence, or under another key (Appendix M.8). Here, an official failure did not always mean nondisclosure. Fictional natural-language profiles use a separate field-level rubric (Appendix M.4).

Scope

Every core exact recovery had the intact secret in the model's input, so it shows disclosure, not inference of an absent secret; C1 separately tests rule-guided composition (Appendix M.3). Deleting every memory recorded as derived from the source removed observed exact recovery in Track B, while artificially hiding one recorded link restored much of it, showing reliance on complete provenance rather than its real-world failure rate (Appendix M.5). Track B wraps Mem0 in our own logging, deletion and provenance code and is not a production agent.

Why this matters

For users, deleting a stored interaction record may not stop information derived from it from resurfacing. For providers of memory-enabled agents, deletion should be judged by whether sensitive information survives in derived memory and can still be retrieved or disclosed, not only by removal of the original record. AfterDeleteBench makes this measurable under stated conditions; it does not estimate prevalence in deployed services.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.