acceptodds
Withdrawn from ICLR 2027

When Memory Hurts: Causal Measurement and Joint Read–Write Control in Long-Horizon Agents

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Persistent memory can improve a language agent on average while causing it to fail on individual decisions. We measure such failures with coupled interventions on memory exposure and write-back, followed by retain–delete tests of the records that exposure induces. Our benchmark, NMR-LOOP, covers embodied control, web workflows, and evolving personal state. On organic streams without injected records, retrieved memory turns a success into a failure—negative memory re- trieval (NMR)—on 4.5% [4.2, 4.8] of paired probes despite a mean benefit of 2.6 percentage points [2.2, 2.9]; brackets denote 95% stream-clustered bootstrap confidence intervals. We then introduce LOOPGUARD, a joint read–write controller trained on intervention labels for immediate utility and for harmful records induced within a finite follow-up window. At comparable positive-transfer rates near 12%, LOOPGUARD attains a controlled-evaluation NMR of 3.3%, compared with 9.5% for dense retrieval, 5.8% for Causal Memory Intervention, 5.3% for MeClear, and 4.7% for its read-only variant. Joint control also moves write-back interaction esti- mates toward zero in all three domains over eight future episodes. At deployment, LOOPGUARD executes each task once, with no additional counterfactual reader calls and a 6.2% token overhead over dense retrieval.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Agent Memory, Causal Evaluation, Closed-Loop Control

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_mFzUOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how an agent’s memory can cause mistakes even when it improves average performance, and how those mistakes can spread when they are stored and reused. The authors introduce NMR-LOOP, a benchmark that compares decisions with and without retrieved memories and tests whether newly stored records harm later tasks. They also propose LoopGuard, which controls both what the agent retrieves and what it saves, using training signals from these interventions. Across tasks, LoopGuard reduces memory-caused failures while preserving similar benefits from useful memories.

Critical strengths
  1. The paper goes beyond average performance to examine when retrieved memories cause failures and whether those failures produce harmful records that affect later decisions.

  2. LoopGuard reduces harmful reversals compared with competing methods and its read-only variant while preserving similar positive-transfer rates, with modest reported deployment token overhead.

Critical weaknesses
  1. The models, controller architecture, and training settings are not clearly specified, making the results difficult to reproduce.

  2. Lower memory-induced harm at similar positive-transfer rates does not necessarily imply better end-to-end performance, since each policy is compared against its own no-memory reference. Absolute task-success results would strengthen the practical claim.

  3. LoopGuard is compared mainly on controlled streams, while the evaluation only measures dense retrieval. An ablation removing descendant-risk prediction is also needed to establish its contribution beyond immediate-utility prediction and write filtering.

Questions for authors

See weakness.

Reviewer_v4HBOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This work proposes NMR-Loop, a benchmark to analyze the effect of outdated memory. Specifically, they consider tasks where that an agent may have some memory due to previously completing it, but this memory may hurt the agent later on if the task changes. Hence, existing memory benchmarks may show that some system do well on average, but hide specific failure modes of this form. The benchmark and evaluation framework test a few different things:

  • An agent’s ability to complete a task with and without the retrieved memory; changes in performance allow for direct causal attribution to that memory
  • The benchmark injects an outdated “seed memory” into the memory, and analyzes the “write-back amplification” whether this seed agent causes the agent to write-back further harmful memories The benchmark covers WebArena and AlfWorld style tasks. Using this benchmark, the authors propose LoopGuard, a trained network that is trained on the benchmark labels to estimate the utility of a memory; it filters out existing memories based on their estimated utility, and new memories that might have harmful consequences in the future. Overall, LoopGuard exhibits lower NMR (the fraction of time at which a memory hurts the agent) than previous methods.
Critical strengths
  • The problem motivation is interesting and useful; understanding how to prevent agents from overrelying on old, outdated memories is a key failure in SOTA memory systems
  • The benchmark is reasonably novel, in particular it studies the issue of “write-back amplification”, how bad memory writes affect downstream memories, which is an important failure mode in memory systems.
Critical weaknesses
  • The paper is difficult to read; terms, algorithm names, and technical jargon are not clearly defined before using making the prose very hard to follow (I.e. the term “seed” is used multiple times before its clear what this means in the context of the benchmark)
  • The main method, LoopGuard, is only trained and evaluated on the benchmark that the work itself proposes; its not clear that the trained model can be effective on OOD tasks, which is important for this to be useful
  • Reproducibility is highly limited; details about which LLM was used, how methods were implemented and tuned, and even the LoopGuard encoder architecture are non-existent
Questions for authors
  • Can the authors provide ablations on the lambdas in the read loss, and the taus used for the write gate thresholds; there are a very substantive number of hyperparameters here and its not clear that the method can be feasibly used without prohibitively substantial hyper parameter tuning?
  • Can the authors show that LoopGuard works on tasks from other benchmarks, and is stronger than the baseline methods?
  • Why are the results shown at a positive-transfer rate of 12%? Can the authors show the tradeoff at a variety of different PMR levels for both their method and baselines? Its unclear whether baselines may perform better at a different PMR
  • Can the authors report network architectures, parameter counts, etc. for baselines and their method?
Reviewer_ik9POpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies how retrieved memory can negatively affect LLM agent performance. It introduces NMR-Loop, a benchmark for measuring harmful memory retrieval and its downstream effects, and LoopGuard, a controller that jointly manages memory retrieval and writing. Experiments across embodied control, web workflows, and evolving personal-state tasks report reductions in harmful retrieval compared with existing baselines.

Critical strengths

The paper addresses an important problem in persistent-memory agents: retrieved information can sometimes hurt performance. Evaluating both immediate retrieval effects and longer-term write-back consequences is a potentially useful direction.

Critical weaknesses
  1. Limited novelty and unclear research gap. The central idea that stale or incorrect memories can harm agent performance is already well understood. Prior work on causal memory selection, memory contamination, and risk-aware retrieval addresses closely related problems. The paper does not convincingly explain what fundamentally new insight or technical contribution LoopGuard provides beyond combining existing techniques.

  2. Insufficient experimental transparency and reproducibility. The paper reports results over thousands of streams, but provides limited implementation details about how the environments, intervention labels, and controller training data are constructed. Important information about the training process, model architectures, and experimental setup is missing. This makes it difficult to independently assess the reported improvements. This section needs significant improvements there.

  3. Questionable generalizability and evaluation design. The controlled experiments focus on memory-optional tasks and predefined harmful-memory scenarios. It is unclear whether the reported improvements transfer to realistic long-horizon agents with more complex memory dependencies. The experiments also do not sufficiently establish that joint read–write control generalizes beyond the particular intervention distributions used for training and evaluation.

Questions for authors
  1. What is the fundamental technical novelty of LoopGuard compared with existing causal memory selection and risk-aware retrieval methods?
  2. Can the authors provide complete implementation and training details, including the controller architecture, training data construction, and intervention-label generation?
  3. How well does the method generalize to unseen memory failure modes and real-world long-horizon tasks beyond the controlled settings?

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

Related papers

Discussion (0)

Sign in to comment.