acceptodds
Withdrawn from ICLR 2027

What Deserves Memory? Behavioral Counterfactuals Make LLM-Agent Memory Self-Evolving

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Any memory-equipped LLM agent must decide which experiences deserve to be kept. Three answers are in use—introspective self-evaluation, retrieval relevance, and LLM-managed abstraction—and none has been compared against the others under controlled conditions. We make the decision measurable by assigning each entry a causal credit: the leave-one-out drop in task performance on a small leak-free probe set, which is exactly that entry's first-order Shapley term. Entries of non-positive credit are evicted before retrieval. We prove that this rule discards an adversarial entry however similar it is to the query, and that it recovers the true causal ranking of memories under stated coverage conditions. Across five domains, baselines, and five backbones from three model families—all same-backbone, paired, and multi-seed—causal credit matches the ground-truth curation oracle, where introspection is worse and correlates with true correctness at . It improves on CHECK on both multi-hop editing datasets, leads all eight memory frameworks on conflicting-fact consolidation and seven state-of-the-art systems on LoCoMo, and is the only defense we tested that combines zero attack success under poisoning with intact benign accuracy, at 9 LLM calls per example—the lowest in its estimator family. Its advantage follows a bank-size law () that also identifies the regime where causal credit is unnecessary.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
LLM agent memory, Causal credit assignment, Memory curation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_5X2jOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies the problem of memory management in LLM agents. The proposed method measures the leave-one-out performance of each entry in the memory and keeps only entries with positive contribution. Results on several datasets are promising.

Critical strengths
  • The paper studies a timely problem. However, the problem becomes less and less relevant as LLM getting more and more powerful.
  • The proposed method, LAMARCK, is simple to implement. However, the problem setting is unrealistic and self-contradictory. More about this in the weakness section.
Critical weaknesses
  • The paper is AI generated with broken language and structure. It is painful to read this kind of paper.
  • The proposed method lacks novelty, which is typical for AI generated papers. Deciding to keep or discard an entry based on performance change on a sealed test set is a very old idea.
  • The problem setting is unrealistic and contradictory. More concretely, the adoption/rejection rule is based on performance on a sealed static test set. In production systems, the environment evolves over time, so a static test set quickly becomes outdated. The authors might argue that their approach is valuable in a static environment, but a static environment doesn't need a memory system in the first place. Furthermore, sometimes the agent has to memorize something that hurts the test set performance because that's what the user likes. The example in the paper's introduction is actually an example of this scenario, and the paper didn't handle it correctly. I doubt that the human authors of this paper even read the manuscript before submitting.
Questions for authors
  • Have you read what the AI generated?
Reviewer_zJaqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes LAMARCK, a framework for curating LLM-agent memory through behavioral counterfactuals. Each memory entry receives a leave-one-out (LOO) credit based on the change in task performance when the entry is removed, and entries with non-positive credit are removed before retrieval. The framework is instantiated for knowledge editing, fact consolidation, dialogue memory, and poisoning defense. The paper reports broad empirical improvements and provides theoretical claims concerning poisoning immunity and causal-order recovery.

Critical strengths
  1. The paper clearly distinguishes retrieval relevance and model-assessed plausibility from a memory entry's actual contribution to task performance. This distinction is particularly valuable when current edits conflict with a model's prior knowledge.

  2. The experiments that hold the backbone, candidate pool, and memory budget fixed while changing the selection signal are persuasive. The removal-order experiment also provides more direct evidence about the mechanism than aggregate benchmark improvements alone.

Critical weaknesses
  1. Equation 1 defines credit as a signed difference in task performance: an entry deserves positive credit only if retaining it improves the outcome. However, Section 3.4 defines per-hop CCC credit as an indicator that removing an entry changes the model’s self-answer. An answer change establishes influence, not benefit. For example, removing a malicious entry could restore the correct answer, yet the answer-flip indicator would still assign that entry positive credit. The paper should clarify whether an additional correctness signal determines the direction of the change and how that signal is obtained. Without this clarification, the CCC implementation does not clearly inherit the interpretation or guarantees of Equation 1.

  2. Proposition 1 establishes eviction when an entry has non-positive contribution on every probe. The most difficult step is exactly showing that malicious entries satisfy this premise. An entry could combine useful information with a conditional malicious instruction, improving ordinary probes while harming a particular target query. Its average credit could remain positive even if the probe set covers that target. Therefore, coverage alone does not establish immunity under the mean-credit rule. The reported zero attack success is encouraging for the tested injection setting, but varying contamination rates and seeds do not test mixed-utility or adaptive attacks. The guarantee and empirical claims should be scoped accordingly.

  3. In Proposition 2, monotonicity does not guarantee agreement in contribution magnitudes or rankings across coalitions. For example, consider three memories a,b,ca,b,c: 78% of queries can be answered using either aa or bb, 1% require aa, 1% require bb, and 20% require cc. Assuming success whenever the required memory is available, this setting is monotone and permits complete probe coverage. The full-bank LOO contributions are 0.01,0.01,0.200.01,0.01,0.20, whereas the Shapley values are 0.40,0.40,0.200.40,0.40,0.20, respectively. Thus, the rankings reverse even though every memory has strictly positive LOO credit and there is no sampling error. Furthermore, the proof appears to move from probe coverage to the stronger assertion that each entry’s contribution is dominated by probes from its own family. Coverage alone does not establish this.

  4. Algorithm 1 computes all credits against the same original bank and then removes every non-positive entry. Consider two interchangeable correct memories: removing either individually leaves performance unchanged, so both receive zero credit, but deleting both destroys the information needed for the task. Thus, individual dispensability does not imply joint dispensability. This is particularly relevant to memory banks containing repeated or paraphrased facts. Sequential rescoring, group-level selection, or an explicit representative-retention rule could address the issue, but these mechanisms are not specified in Algorithm 1. The relationship between its threshold rule and the one-entry-per-family budget also needs clarification.

Questions for authors

refer to weaknesses

Reviewer_gHgdOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes to filter an external memory bank by introducing a method named causal credit. This method considers a probe set of questions related to the target task. For each entry of the memory bank, the overall performance on the probe set is evaluated with and without that entry (leave-one-out of the probe set), and the entry is kept only if the difference (called credit) is greater than a positive threshold.

Two properties are highlighted: poisoning immunity (against a single entry that has a harmful impact on the probe set) and causal-order recovery (claiming that the credit score recovers the true importance ordering, under some assumptions).

The method is tested on classic benchmarks (Tab. 2: MQuAKE, LoCoMo, etc.) and some unspecified benchmarks (e.g. "scaling suite" in Tab. 2), for a total of 8 benchmarks. The model used in the main paper is GLM5.2. One highlighted result is a x9.1 improvement compared to the "introspection" method. Many different experiments are provided in Section 5 and in the appendix.

Critical strengths
  • I think that the proposed filtering strategy can help in improving the benchmark performance, after carefully building the probe set.
Critical weaknesses
  • The quality of the presentation is low. Some terms are not well defined, or not introduced at all. Importantly, the probe set is critical but only explained in l.196 (Leak-free protocol) without further explanation.
  • There are too many experiments, in many directions. Some parameters might have been fit post-hoc (e.g. the % of facts removed in Fig. 3, where the maximum in Fig. 3(middle) is exactly the performance reported in Fig. 3 (left))
  • I don't think the idea can survive under a real benchmark.
Questions for authors

The questions follow the reading of the paper, and are related mainly to the presentation and to the methodology.

  • The term "introspection" in the abstract is never defined in the paper. In l.106 it is related to the reference Singh et al., but this paper is a "reality check" and does not introduce new methods. My guess was that introspection meant from parametric weights. If so, the model obviously does not know the edited facts and cannot answer correctly, and the "introspection is 9.1x worse" is not surprising (for reference, x9.1 means from 0.03 [with the introspection method] to 0.30 [using the Lamarck method proposed in the paper]). Overall, putting it as the main comparison point is not a strong signal of performance improvement for the proposed method.
  • What are the compared elements for the "correlation"? Assuming my previous guess is correct, showing a strong negative correlation when asking about an edited facts is normal (This comment is related to the sentence "correlates with true correctness at -0.82").
  • The "bank-size law" in the abstract is deduced from 3 points in Fig.5b. I don't understand, on that figure, what is the model, the task, etc. (the caption only says: "Advantage vs. bank size, log2 fit, R2=0.997", and the main text l.418 only state: "Controlled synthetic sweeps". The term "synthetic" is only present further in Table 2 (the row: "scaling suite, controlled synthetic banks" with metric "advantage" and n=72 cells). There is no further information in the text about it.
  • The Table 2 is compact and confusing.
  • In the intro l.34, "a planted note" is not precise. The introduction might use high-level language, but I don't understand whether you are referring to parametric weights, external memory, or in context memory. Globally, I guessed that: March=parametric, June=external and Building7=in-context.
  • In l.45 the authors state "this paper supplies that comparison (Fig. 2)" but Fig.2 labels only contain "introspection".
  • I agree with lines 47-49. Accumulation of contents in external memory banks is harmful if incorrectly managed.
  • B=1 is undefined in l.100.
  • The table 1 is not understandable at that stage.
  • l.146--152 are difficult to read
  • In l.57, the whole methodology consists in discarding non beneficial entries. However, a memory that is not useful at time T might be useful at time T+1. In the tested benchmarks, it might not be the case, but generally, asking "In which building does the user move in June?" at T+1 might be un-answerable if this information has been deleted previously. Overall, I think that filtering might give some help, but only after checking the contextual validity of each assertion (each sentence presented in l.33-34 is correct when expressed in the right context).
  • The training/test prevention of contamination is detailed in l.196-199 (describing how the credit Eq. 1 is computed practically), but the whole protocol is not clear. Can you provide at least a couple of examples of your methodology?
  • l.210 proposition 1. The term Delta_e is not defined (only defined in l.722).
  • The algorithm 1 describes one selection round (driven by Eq. 1), but the number of episodes is unclear. I only realized that the number of facts removed can be large in Fig.3 (middle), where 50% of the facts were removed. Actually, how this number "50%" has been obtained? Is the performance maximized over the "facts removed" parameters. In Fig. 3, the 0.3 performance is exactly the performance maximizing the lower-credit first curve in Fig. 3 middle.
  • The threshold tau (only defined in Alg. 1) is linked with the % of facts removed by the correspondence is not 1to1.
  • Regarding proposition 1, the assumption (ii) is never defined (in the proof shown in the appendix). In particular, this assumption is not mentioned in the main.
  • There is an issue in removed recursively the elements by batch of 1. Assume that the memory bank is poisoned with 2 almost identical elements "level 7 always". Then the performance when removing one element is identical, and neither poisoned e1 and e2 are removed.
  • In l.262, "Published numbers from other papers [...] are shown only as reference, never compared". Where are those reference results?
  • l.266-269 appear to be fully LLM generated. Same from 283 to 294. Sec. 5.2 is LLM generated and does not add much for a human
  • l.309, why causal credit methodology reaches only 0.3, and why does gt oracle only reach 0.3? Can you provide an example of a positive and of a negative answer?
  • What is "keep-all" (I assume this is in-context learning), and why in Fig.3a the "keep-all" reach only 0.12? I checked that GLM5.2 allows 1M context size, what is the problem in ingesting one full dialogue of LoCoMo in-context
  • How "Curation itself stayed perfect (1.000) in every cell", and how to justify exactly perfect performance?
  • In the conclusion, I don't understand the sentence: "the reason is structural rather than empirical". Can you justify it?
  • In l.482, the authors state: "causal credit loses where the pool already fits in context", but this is not documented in the main paper. Maybe the authors are referring to l.1118?

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

Related papers

Discussion (0)

Sign in to comment.