acceptodds
Under review as a conference paper at ICLR 2027

Amortizing Causal Memory Verification via Evidence Reuse in Long-Horizon LLM Agents

Abstract

Long-horizon large language model (LLM)-based agents increasingly rely on external memory, but relevance-based retrieval does not establish whether memory exposure improves agent behavior. Causal Memory Intervention (CMI) provides a stronger behavioral criterion through matched interventions, but applying Full-CMI to every candidate incurs repeated verification costs. We study whether this cost can be amortized across sequential memory decisions. We propose the Causal Evaluation Reuse Policy (CERP), a selective verification policy that fits a rejector with history-conditioned features on an initial Full-CMI-evaluated prefix and applies the frozen policy to subsequent candidates, rejecting candidates before utility verification. On Causal-LoCoMo, CERP achieves 12.7% net call savings relative to Always-Verify after accounting for the initial Full-CMI cost, with a pooled rejection error rate of 8.8%, a 0.02 decrease in the extractive task proxy, and a 0.051 decrease in Useful-Memory F1. Matched random rejection and direct nearest-label reuse incur higher rejection error (16.3-23.7% vs.8.8%). Budget and horizon analyses characterize when the initial verification cost is recovered.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Long-Horizon LLM Agents, Agent Memory, Causal Memory Intervention, Cost Amortization

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_suY2OpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper explores how to reduce the memory verification cost of long-term LLM intelligent agents. The author proposes CERP: Firstly, perform Full-CMI on a small number of candidate memories, and use the verification results to train a lightweight classifier that combines historical evaluation information with the similarity to the current task; subsequently, reject the expected useless memories in advance, and only perform cheaper utility verification for the remaining candidates, thereby amortizing the initial cost. The paper has been verified on Causal-LoCoMo. CERP saves the number of calls while resulting in a decrease in the Useful-Memory F1 score and an 8.8% reduction in the error rate of early rejection.

Critical strengths
  1. The paper focuses on the often overlooked verification cost in memory-enhanced agents: Even though verifying each item can improve memory selection, the additional calls may undermine system efficiency. This issue is framed as the cost allocation across sequences for decision-making, providing a useful perspective for evaluating memory verification methods.
  2. CERP employs a small number of historical features and logistic regression to achieve early rejection, with a simple structure. The paper explicitly includes the initial Full-CMI supervision cost and derives the net savings condition NR>BN_R > B, distinguishing between the reduction in the number of calls for relative complete validation and the net gain from relative utility validation.
  3. The experiment separately examined the effects of current similarity, historical information, and their combination, and set a random baseline with the same number of rejections. Under the same call cost, CERP reduced the rejection error rate from 23.7% of the random strategy to 8.8%, indicating that the benefits come from selectively rejecting candidates rather than simply reducing the number of verifications.
  4. The paper analyzed the influence of the initial supervision budget, sequence length, rejection threshold, candidate order, and different responders on the results, and provided validation on a second benchmark. The authors clearly reported a decrease in quality, the inability to recover some orders, and that the utility labels could achieve similar performance, making the contribution boundaries more clear.
Critical weaknesses
  1. Using only the utility verification label can also achieve similar performance. This makes the value of performing additional perturbation verification during the initial stage less clear: if cheaper supervision can train an equally effective rejector, then is the additional cost of Full-CMI worth it?
  2. The main experiments were conducted by replaying the sequences while keeping the candidate set fixed and pre-calculating the verification results. This approach failed to reflect how memory selection affects current behavior, subsequent tasks, and future memories. The task scores of Causal-LoCoMo were derived from the concatenation of retained memories rather than the actual generation of responses.
  3. The main experiment saved 12.7% of the calls compared to Always-Verify, but the F1 score on Useful-Memory decreased by approximately 0.051, and it achieved positive returns only in 7 out of 10 sequences. When the initial budget was increased to B=100B=100, the average net savings became -10.5%. These results indicate that the method has conditional cost advantages, but the stable and predictable deployment benefits have not yet been established.
  4. Classifier fitting and threshold selection Use the same batch of initial samples. The 2% rejection error target set during the training stage does not guarantee the subsequent risk. The actual RER of the main experiment reached 8.8%. The model and threshold were then fixed, while the historical features and task distribution might change; the candidates that were rejected prematurely did not generate feedback, thus lacking a mechanism to discover and correct persistent false rejections.
Questions for authors

See weaknesses.

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

The paper studies whether the cost of Causal Memory Intervention (i.e. CMI) verification can be amortized over sequential memory decisions. Full-CMI evaluates each retrieved memory candidate with matched executions,and then each candidate is retained only when exposure improves the outcome and the improvement is stable under perturbation, therefore it costs two responder calls. The proposed CERP fits a logistic-regression rejector on an initial Full-CMI-evaluated prefix using three history-conditioned features, then routes later candidates: reject before verification, or pay one utility-verification call. Finally some experiments are shown and analyzed.

Critical strengths

The amortization formulation is clean and the accounting is fully transparent. The break-even condition is a clear and auditable identity, and the budget/horizon/operating-point analyses map exactly where amortization is positive, marginal, or negative. This is an useful framing for any expensive per-candidate verification procedure. The experimental setup is detailed. Matched random rejection shows the savings come from rejecting the right candidates, not from rejecting many. The 1-NN and 5-NN label-reuse baselines show that a learned rejector over history features is not equivalent to copying labels.

Critical weaknesses

W1. The utility-only ablation shows that labels skipping the perturbation condition perform comparably to Full-CMI labels. It may result in: 1) the "expensive gold standard" whose cost motivates the whole paper may be replaceable by utility verification at half the cost, so what remains to amortize is then unclear and the title's "Causal" is not established as useful anywhere in the paper. 2) the paper does not test where Full-CMI labels would matter. The two supervision sources are compared only on Causal-LoCoMo, an extractive-QA benchmark. The paper does not report the prevalence of utility-positive but stability-negative candidates (the cases where the two labels can differ) on this benchmark. Nor does it test any regime where placebo effects should be stronger.

W2. Savings are 12.7% relative to Always-Verify, the cheaper half of Full-CMI. Positive amortization occurs in only 7 of 10 replay orderings; roughly 30% of deployments would lose calls. And its quality cost is deterministic and significant (F1 −0.051, CI excludes zero), while the savings are probabilistic at the ordering level. So The net benefit appears modest, and asymmetric in risk. the paper did not convert either side into absolute cost terms, so the reader cannot judge whether the trade is worth taking.

W3. Full-CMI outcomes are precomputed and fixed; replay only permutes evaluation order. In deployment, a rejected candidate never enters the memory bank. This changes the candidate distribution the rejector later faces. And rejected candidates contribute no feedback, so rejection errors are never corrected — a self-reinforcing loop with no online correction mechanism. The longer the horizon, the more both drifts accumulate — yet the horizon is precisely where the paper claims the largest savings. So the offline evidence is least trustworthy exactly where the claimed benefit is largest, and no on-policy experiment verifies that the offline 8.8% RER holds online. The no-rejection fallback concedes the point: the safest response to drift is to stop using the method.

Questions for authors

Q1. Please add a zero-cost self-assessment baseline: ask the responder whether exposing the candidate memory would help, and reject on low confidence. If an LLM judge reaches a comparable rejection error, the learned rejector is unnecessary; if CERP wins, the method's contribution will become concrete.

Q2. The savings are probabilistic at the ordering level (7/10 positive) while the quality cost is deterministic. Can you convert both sides into absolute cost terms for a realistic deployment — cost per responder call, cost of F1 degradation? And can you report the per-deployment variance of net savings alongside the mean?

Q3. The offline replay fixes all Full-CMI outcomes, so rejection errors are never corrected online. Have you tried the minimal feedback variant — spot-checking a small fraction of rejected candidates and adding them to the rejector's training set? If online RER drifts away from the offline 8.8% in even a small live-responder experiment, the 12.7% figure should be treated as an upper bound.

Reviewer_9dBfOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Causal Memory Intervention (CMI) is a criterion for deciding whether a retrieved memory should be exposed to an LLM agent. This paper studies whether expensive memory verification can be amortized across sequential decisions and proposes Causal Evaluation Reuse Policy (CERP). CERP initially requires Full-CMI labels for a prefix of candidates. It trains a logistic regression model using historical support, historical task similarity, and current memory–task similarity. Subsequent candidates are either rejected immediately or passed to utility verification. On Causal-LoCoMo, CERP reduces candidate-level calls from 500 to 436.3 (12.7% net savings) with 8.8% rejection error, a task-proxy decrease from 0.9029 to 0.8829, and Useful-Memory F1 decreasing from 0.8905 to 0.8398.

Critical strengths
  1.  The paper brings selective-prediction and cascade ideas to intervention-based memory verification. It conditions on accumulated evaluation history rather than per-instance confidence.
  2. The cost of verifying retrieved experience is a real obstacle for memory-augmented agents. The proposed small routing model offers an affordable intervention and reducing costs.
Critical weaknesses
  1. CERP's losses vs. Always-Verify (proxy −0.020, F1 −0.051) exceed Full-CMI's entire advantage over Always-Verify (+0.003 and +0.046).
  2. The paper never says how the verification signal would exist at inference time. If it needs the current task's gold answer, verification cannot run before the task is answered and an oracle scorer could simply select outputs directly.
  3. The evaluation does not test the claimed setting of long-horizon LLM agents and also in every main-text result, no LLM acts on the retained memories.
  4. Fig. 1 motivates reusing historical outcomes, but the rejector does not use outcomes as features. In Causal-LoCoMo, useful memories are written separately for each task while distractors are sampled from the same conversation, so recurring memories are likely distractors. CausalMemBench is generated from templates, so smaxs^{max} may simply detect repeated templates.
Questions for authors
  1. Why does Full-CMI needs amortizing?
  2. On the high-disagreement subsets, does CERP preserve Full-CMI's advantage over Always-Verify at lower cost? What are its false-acceptance and poison-adoption rates?
  3. Please add risk–coverage curves for similarity-only, history-only, and full CERP and also a cost–quality frontier that includes Always-Verify.
  4. What is s(⋅,xt)s(·, x_t) at inference time, and does it require the current task's gold answer?
  5. Does a policy fit on Causal-LoCoMo transfer to CausalMemBench, and vice versa?
  6. How does a frozen policy behave on a longer conversations with memory writes and topic drift?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.