acceptodds
Under review as a conference paper at ICLR 2027

The Seven-Year Itch of Lifelong Agents: When Perfect Memory Prevents Adaptation

Abstract

Long-lived AI agents rely on memory to personalize over time, yet the user they remember may change. An assistant can accurately retain years of past preferences and still make poor decisions for the user now if obsolete evidence continues to shape current responses. We call this crossover the seven-year itch of lifelong agents: beyond some interaction history, retaining more behaviorally active memory can reduce current-state adaptation. We characterize the effect with a drifting-user model that separates the benefit of averaging more observations from the bias created by stale evidence. This analysis motivates (Finitude-Aware Memory), which preserves historical episodes while adapting how strongly they influence current personalization; commitments follow a protected path, and a compact handoff state carries forward still-valid information. We evaluate the resulting design with , a controlled benchmark spanning eight forms of user change, six LLM backbones, and repeated runs. The study reframes long-term memory as a temporal-validity problem: the challenge is not only to remember the past, but to decide how much authority the past should retain over the present.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
lifelong agents, long-term memory, user adaptation, non-stationary personalization

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_vi3COpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates a failure mode of lifelong personalized agents: retaining historically accurate memories can hinder adaptation when user preferences change. It introduces ITCHBENCH, a benchmark covering stationary and nonstationary preference trajectories, and proposes Finitude-Aware Memory (FAM), which separates archival retention from the influence of memories on current decisions. FAM combines learned validity estimates with relevance-based retrieval, protects commitments and constraints, and constructs bounded states for memory handoff. The paper also provides a theoretical analysis of the trade-off between estimation variance and stale-information bias. Experiments across six language-model backbones report improved adaptation and a favorable balance between plasticity and continuity.

Critical strengths

The paper addresses an important problem in long-term agent memory: historically accurate information can become misleading when user preferences change. Evaluating adaptation alongside continuity, including transient and context-dependent changes, provides a useful perspective on memory management.

Critical weaknesses
  1. Insufficient reproducibility of the benchmark and method. ITCHBENCH is a central contribution, yet the materials available for review provide no accessible release of the dataset, generator, evaluation code, or FAM implementation. Although Appendices B–C describe the framework, essential details remain insufficiently specified, including trajectory-level data partitioning, the controller’s training objective and optimization settings, feature extraction, and the construction and scoring of evaluation probes. The reference to results/drift_family_summary.csv does not resolve this issue because the file itself is unavailable. These omissions substantially limit independent verification of the benchmark and reported improvements. An accessible implementation or a sufficiently complete specification is needed.
  2. The interpretation of the theoretical calibration requires clarification. Theorem 2 provides a finite-window risk upper bound, and Appendix A derives the retention rule by minimizing that bound. However, Figure 3c and Table 5 compare the resulting scaling exponents with those of the empirically optimal lifetime. The upper bound alone does not establish this correspondence. The authors should clarify the process assumptions and interpretation of the calibration, and distinguish the proposed tuning rule from a claim about the actual optimal window across different drift processes.
Questions for authors
  1. Section 4 states that methods use a matched 4,096-token active-memory budget, whereas Table 13 reports 24,500 tokens for full history. How should these figures be reconciled, particularly given the matching accuracy results in Tables 2 and 13? Does the budget include protected information and handoff states?
  2. In Figure 3b, does introducing stale information also remove current evidence? Can current evidence and total prompt length be held fixed while comparing obsolete contradictory statements, neutral distractors, and explicitly time-qualified historical facts? This would help distinguish stale-memory interference from the effect of reduced current evidence.
  3. Which controller inputs, such as attribute identifiers, contextual scopes, confidence estimates, and protection flags—are inferred from dialogue, and which are supplied by the benchmark’s latent annotations? Are all three renderings of each trajectory assigned to the same split, with future-consistency supervision restricted to training trajectories?
  4. Since Equation 5 makes an episode’s validity non-increasing, how can retired memories become usable again during cyclic changes or preference reversals? What precisely constitutes “renewal” and “rollback” in Table 12, and how are baseline handoff states constructed under comparable capacity constraints?
Reviewer_1cZtOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies personalization when a user's preferences change but earlier observations remain in memory. A weighted-estimation model separates stale-state bias from observation noise. Motivated by this trade-off, FAM uses a learned validity controller to reduce the retrieval weight of older episodes while preserving an archive, protected constraints, and a compact handoff state. ITCHBENCH generates eight types of latent user trajectories and renders them as dialogue. The paper reports improved adaptation across six answer backbones, together with continuity, handoff, robustness, and component-ablation results.

Critical strengths

The problem is valid and interesting, especially for long-lived agents interacting with users whose preferences change over time. The proposed approach of estimating temporal validity to manage memory influence is reasonable, and the authors provide extensive empirical evaluations on a controlled synthetic benchmark.

Critical weaknesses
  1. Limited theoretical novelty. The mathematical analysis largely follows the standard bias-variance trade-off under non-stationarity and does not offer particularly surprising insights. This is not a major concern if the theory is primarily intended to motivate the algorithm rather than serve as a key contribution.

  2. Insufficient comparison with existing methods. The paper does not clearly distinguish FAM from existing approaches to adaptive memory management and temporal validity, such as MemoryBank, Zep/Graphiti, Memory-R1, and AgeMem. Although A-Mem and several standard baselines are evaluated, comparisons with more closely related methods would help establish the novelty and advantages of FAM.

  3. Limited benchmark generalizability. The experiments are primarily conducted on ITCHBench, which is designed by the authors around the same temporal-validity problem that FAM addresses. This raises concerns about whether the benchmark particularly favors the proposed approach. While the authors include robustness and held-out evaluations, these remain within the same synthetic framework. Evaluations on independent benchmarks, such as LongMemEval or LoCoMo, would make the empirical evidence more convincing.

Questions for authors

See Weaknesses section

Reviewer_i1KfOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates how outdated memories continue to influence long-term personalized agents after users’ states change. It refers to the phenomenon in which current-state accuracy declines as interaction history grows as the “seven-year itch.” The authors propose FAM, which learns memory-validity weights to separate archival retention from memories’ influence on current behavior, alongside protected constraints and a compact state summary for memory management. They further introduce ITCHBench to evaluate adaptation, continuity, and state handoff across eight types of user-state trajectories and six answer models.

Critical strengths

The paper addresses long-term memory management when user behavior changes, a problem with practical relevance.

Critical weaknesses
  1. The authors do not discuss how this work differs from prior research. LongMemEval[1] explicitly evaluates updates to user information, while PersonaMem[2] directly studies dynamic user profiling and preference tracking. Although the authors distinguish archival retention from memories’ influence on current behavior, their analysis primarily relies on a simplified setting that estimates user states by weighting historical behavioral evidence. This analysis does not readily generalize to the broader problem of long-term memory retention and selection. What new problem does this paper investigate beyond prior work? What new claims does it make? As presented, its novelty appears limited to terminology and framing.

  2. Long-term memory is a widely studied area with a relatively mature research foundation, yet the paper lacks comparisons with relevant methods. For example, Mem0[3] already marks outdated relationships as invalid while preserving historical records.

  3. Many critical details of the method and evaluation pipeline are missing, and the current description is insufficient for reproduction. The paper does not adequately explain how attribute keys, contradiction features, posterior uncertainty, independent support, and protected bits are obtained from raw conversations. The controller’s specific training loss, the construction of supervision examples, and the data splits are also insufficiently specified. Furthermore, the mapping from natural-language answers to latent-state labels and the scoring procedures for continuity, plasticity, and RSS lack sufficiently precise, operational descriptions.

[1] LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory. ICLR, 2025.

[2] Know Me, Respond to Me: Benchmarking LLMs for Dynamic User Profiling and Personalized Responses at Scale. COLM, 2025.

[3] Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory. ECAI, 2025.

Questions for authors

see weaknesses

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.