acceptodds
Under review as a conference paper at ICLR 2027

Think Before You Remember: Uncertainty-Gated Relational Memory for Long-Horizon Agents

Abstract

Long-horizon agents rely on memory to guide future interactions, but what happens when that memory treats an uncertain judgment as an established fact? In relational memory, an ambiguous observation can prematurely change an agent's trust in another actor, turning a tentative interpretation into a persistent premise for decisions. Unlike an isolated reasoning error, this mistake can persist beyond the interaction that produced it. Storing more history or re-reading past interactions does not by itself prevent this failure; the memory must also decide which inferred changes are justified. We propose Confidence-Aware Relational Memory (CARE), an uncertainty-aware relational memory framework built on a simple principle: remember the evidence without prematurely settling its meaning. CARE separates supported relationship updates from unresolved interpretations. It comprises three modules: (1) Consensus Write applies effects shared across plausible interpretations while gating interpretation-specific updates on sufficient support; (2) Hypothesis Buffer retains unresolved interpretations and their potential effects alongside the original events; (3) Evidence Replay uses new evidence linked to pending events and integrates the supported stored effects once the ambiguity is resolved. Our experiments show that CARE reduces relational-state displacement at ambiguous events by 10.6–46.9% compared with immediate updating. Ablation studies confirm the complementary roles of Consensus Write, Hypothesis Buffer, and Evidence Replay. These findings highlight the importance of controlling what enters relational memory while preserving a path for later revision.

Primary area:
foundation or frontier models, including LLMs
Keywords:
long-horizon agents, relational memory, agent memory, uncertainty, belief updating, social agents, state representation, large language models

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_XvPyOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper claims to identify a problem with long-horizon agents - that they commit to a decision before the evidence supports it. They propose a new method called CARE to fix this problem. CARE consists of three main pieces: Consensus Write (handles the initial write), Hypothesis Buffer (holds the unresolved pieces), and Evidence Replay (when new evidence resolves something previously unresolved, this handles it). This is realized through keeping a four dimensional vector with entries for trust, respect, dependency, and conflict, along with a set of pending records. An LLM gives scores for the entries. They threshold to decide if an event is unresolved or not. They introduce a metric, focal displacement, as a way of measuring the changes when the decision is unresolved.

Critical strengths

I would like to preface this by saying that this is not my field of expertise.

(1) The method is new, although similar to others they cite. For example, BeliefMem seems very similar to me, but they don't ever "commit", rather just update probabilities. This method does commit. To me, it does seem better to commit after a certain point, as providing more information even when it's very unlikely seems to be wasting context.

(2) The experiments are evaluated on a large amount of models (in Table 2 they have Claude, DeepSeek, Gemini, GPT, Kimi, and Qwen), indicating that their gain in focal displacement, relative to the DRSM baseline, holds generally.

Critical weaknesses

I would like to preface this by saying that this is not my field of expertise. I have two problems with the main contribution.

(1) The paper says "We identify the failure of committing interpretation-dependent relational effects before the evidence supports them, and introduce focal displacement as a method-neutral measure of state changes at unresolved events". To my knowledge, I do not think identifying that failure is novel. The paper cites BeliefMem, saying that it "preserves alternative conclusions". I agree with this, but the BeliefMem paper in their abstract says that

"By committing to one conclusion and discarding uncertainty, these methods introduce self-reinforcing error: the agent acts on the stored conclusion, never revisits alternatives, and reinforces the conclusion over time. To address this issue, we propose BeliefMem, which shifts the memory paradigm from committing to a single conclusion per observation to retaining multiple candidate conclusions with their probabilities."

It seems that BeliefMem is identifying this failure. I think that their method/solution is different as they never commit, just keep the several options with a probability associated with them, but they do this because they identified the committing before evidence problem.

(2) To the second part of the main contribution, I am not convinced that the metric introduced, focal displacement, is a useful metric. If you're introducing focal displacement as the metric, I think that you need to make a case for it. The two papers cited as related in the introduction use accuracy, F1, BLEU, SR, etc. as metrics they are evaluating against. As other papers in the field use metrics that are scored against a ground truth or downstream outcome, you need to say why your new metric is valid.

To that point, if my understanding is correct, a method that never writes anything would minimize focal displacement.

Questions for authors

(1) Could you clarify if I am misunderstanding something about your main contribution? The authors say that one of their main contributions is identifying "the failure of committing interpretation-dependent relational effects before the evidence supports them". It seems to me that BeliefMem (which is cited in the introduction), and some others, make this identification. Could you explain the difference between what you and they identify?

(2) Could you explain or provide experiment evidence for why focal displacement is a good metric? If my understanding is correct, a method that never writes anything would minimize focal displacement. Without other established metrics accompanying your results, I don't know how to interpret your results. This is my main issue with the paper.

(3) As a smaller point, the paper seems really intent on telling me what experiment was done with the "historical" vs "revised" implementations, first introduced in footnote 1, although this is never defined. Most of the experimental figures and tables in the main paper (all except Figure 5 I believe) use the historical implementation, whatever this means. If this really matters enough to warrant a distinction in the main paper, then shouldn't the experiments be re-run with the revised implementation, or at least have the paper tell us the difference?

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

This paper focuses on a dilemma in long-horizon agents, where an uncertain judgement is treated as an established fact in memory, and the memory is further used to guide the interactions. Meanwhile, in those relational memories, an ambiguous observation can prematurely change an agent's trust in another actor, turning a tentative interpretation into a persistent premise for decisions. What's worse, strategies like storing more histories or rereading past interactions can hardly help. Thus, in this paper, a method, Confidence-Aware Relational Memory, is proposed based on a simple principle: remembering the evidence without prematurely settling its meaning. Empirical results show that the proposed method can effectively solve the problem mentioned above.

Critical strengths
  • The motivation of the paper looks good. With the developments of advanced LLMs, human beings expect these powerful foundation models to perform more complicated tasks. Due to the complexity of these tasks, it takes much longer for reasoning. In this context, the reliability of the memories is essential for the final reasoning results.

  • The principle "remember the evidence without prematurely settling its meaning" proposed in this work is a simple yet meaningful case. This principle further narrows the scope of unreliable memories to the overthinking of LLMs. It is essential to distinguish the cases between retaining the existing evidence and adding necessary interpretation.

  • This paper honestly reports the negative results and limitations, which provide an authentic view of the proposed method.

Critical weaknesses
  1. The metric adopted in the experiments cannot support CARE producing better judgement. In the experiments, focal displacement is adopted for measurement. However, lower focal displacement only measures restraint, while it cannot reflect the correctness. In a case where a policy is never updated, a zero displacement can also be achieved. Moreover, ambiguous evidence can still justify a meaningful change in trust. Thus, the reduction does not demonstrate the advantage of CARE, unless comparisons are made to simple policies, such as smaller updates.

  2. In this paper, commitments are determined by uncalibrated support scores from LLMs and predefined thresholds, and the "shared effect" is adopted according to the evaluation of the same LLM. In this case, the consensus among generated interpretations can hardly demonstrate that an effect is justified due to the unstable and unreliable self-evaluations (cf. Line 174). Moreover, since only a unique optimal candidate is selected, there exists a case that a candidate wins among all candidates with a very small advantage in scores (cf. Line 185-186).

  3. In the proposed CARE method, interpretation-specific effects are frozen when the original event happens. These will then be used without evaluating their magnitude with changed states (cf. 205-206). Moreover, the initially committed events are also unavailable for replay. These aspects somehow prevent the proposed mechanism from reversing the earlier confident mistakes.

  4. In the experimental setting, ambiguous events and clarification links are available, histories are controlled by programs, and reference states are generated from LLMs. With these aspects taken into consideration, the experiments cannot validate whether an agent can identify ambiguity, build connection to later evidence, or improve decisions.

Questions for authors
  1. Have you ever considered adopting updates that are simple and smaller or delayed as baselines? In such a case, can CARE outperform these baselines on state correctness and downstream decisions?

  2. Is the commitment gate robust enough in cases where interpretations are missing, the support scores are inaccurate, and the thresholds vary?

Reviewer_3TEnOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses premature commitment in relational memory for long-horizon social agents: when ambiguous events cause interpretation-specific state updates that persist and misguide future decisions. The authors propose Confidence-Aware Relational Memory (CARE), a framework consisting of three components: Consensus Write, Hypothesis Buffer, and Evidence Replay. The core idea is to separate shared effects across plausible interpretations from interpretation-specific effects, committing the latter only when sufficiently supported while retaining unresolved interpretations for later revision. Experiments across six model families and controlled “paired worlds” suggest CARE reduces early state displacement under ambiguity while exposing that robust post-clarification revision remains challenging, particularly for malicious clarifications.

Critical strengths
  1. The paper addresses the practical issue of premature commitment in long-horizon agents.
  2. Figures 1 and 2, along with Algorithm 1, clearly illustrate the motivation and proposed method.
  3. CARE combines three complementary modules to manage uncertainty and enable subsequent memory revision.
Critical weaknesses
  1. CARE relies on uncalibrated LLM-generated support scores and fixed gating thresholds (τplaus),(τdom)(\tau_{\text{plaus}}), (\tau_{\text{dom}}), with limited sensitivity analysis. This raises concerns about its robustness across different models and scenarios.
  2. Candidate interpretations, shared/specific effect decomposition, and evaluation reference intervals all depend on LLM judgments. Their reliability is insufficiently validated, making it difficult to distinguish improvements in the memory mechanism from the quality of LLM-based assessments.
  3. CARE substantially underperforms DRSM under malicious clarifications. Although the paper acknowledges this limitation, it lacks a detailed analysis of the underlying causes, raising concerns about the effectiveness of Evidence Replay.
Questions for authors
  1. How sensitive is CARE to the gating thresholds ((τplaus),(τdom))((\tau_{\text{plaus}}), (\tau_{\text{dom}})) and the number of candidate interpretations? Could the authors provide sensitivity analyses or evaluate calibrated support scores?
  2. How is the reliability of LLM-generated candidate interpretations and shared/specific effect decomposition validated? How would errors in these assessments affect CARE's performance?
  3. What causes CARE's substantial performance degradation under malicious clarifications? Is it related to inaccurate stored effects, support-score updates, or the Evidence Replay mechanism?
  4. Can the authors evaluate CARE on realistic end-to-end long-horizon agent tasks to demonstrate improvements in memory accuracy and downstream decision-making?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.