acceptodds
Under review as a conference paper at ICLR 2027

PraMem: Practice-derived Experiential Memory for Long-horizon Behavior Prediction

Abstract

Long-horizon behavior prediction plays a crucial role in artificial intelligence application, which aims to infer a user's next action based on a lengthy historical sequence. The rise of large language models (LLMs) offers a promising direction for sequential behavior prediction, but LLMs still struggle with latent behavioral pattern induction and model-intrinsic cognitive biases when tackling long-horizon behavior prediction. Prior memory management methods follow a context-compression paradigm that attempts to address this task by alleviating the historical sequence burden, yet fail to resolve core challenges. In this paper, we advocate a paradigm shift that reframes the lengthy historical sequence from a burden into a valuable resource to be exploited, and accordingly propose PraMem, which conducts practice over lengthy historical sequence to build an experiential memory, thereby serving as assisted input for accurate long-horizon behavior prediction. Extensive experiments across diverse tasks demonstrate PraMem achieves superior performance than prior methods, and more in-depth analyses provide valuable insights into mechanism of the experiential memory.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Experiential Memory, Long-horizon Behavior Prediction

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_YRm7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PraMem, which builds experiential memory consisting of user behavioral patterns and bias alerts for the model through prediction practice, reflection, and periodic updates over users’ historical behavior sequences. The method further introduces groundedness review based on historical sequence perturbation and generalizability review based on scene identification to filter proposals for memory updates. Experiments on three datasets show that the method outperforms the compared baselines in most scenarios.

Critical strengths
  1. Leveraging ground-truth feedback from historical behaviors to construct experiential memory is well motivated. It directly connects memory updates to the prediction objective without updating the underlying model parameters.
  2. The experiments cover three datasets and two prediction backbones, with notable accuracy improvements in scenarios such as advertising and live streaming.
  3. The paper provides ablations of the core modules and an analysis of memory evolution, and demonstrates that the constructed memory can be reused by another prediction model, offering preliminary evidence of its practical utility.
Critical weaknesses
  1. The groundedness review does not adequately distinguish perturbation sensitivity from groundedness in historical evidence. A bias-alert experience supported by repeated prediction errors may remain valid even when the user's behavioral pattern changes substantially. Its invariance to such perturbations does not necessarily imply a lack of historical grounding. Conversely, an unsupported proposal may be judged invalid after perturbation and thus pass the groundedness review.
  2. The generalizability review may conflate a lack of specificity with generalizability. Equation (7) assesses generalizability based on whether a proposal uniquely identifies the original scene. However, an overly general statement that cannot identify the original scene may have no predictive value, whereas a valid brand-preference experience may be rejected simply because the original scene is the only candidate involving that brand.
  3. The analyses in Table 3 and Figure 5 do not fully validate the proposed self-review criteria. The audit sample size, annotation criteria, and annotator agreement are not adequately described, and the precision and recall of self-review are not broken down by review criterion or experience type. Figure 5 supports the usefulness of the modules, but does not directly establish that their acceptance decisions reliably reflect groundedness and generalizability.
  4. The paper should clarify its novelty over existing feedback-driven methods. Reflexion [1] already uses task feedback to construct reflective memory, while DGDPO [2] diagnoses and revises user profiles based on discrepancies between simulated and actual user behaviors. Although Section 5 highlights bias-alert experience and self-review as distinctions, these differences do not establish a paradigm shift from context compression to practice-derived memory. A more focused comparison with feedback-driven approaches would help clarify the contribution.

[1] Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, Shunyu Yao. Reflexion: Language Agents with Verbal Reinforcement Learning. NeurIPS 2023. [2] Hongyang Liu, Zhu Sun, Tianjun Wei, Yan Wang, Jiajie Zhu, Xinghua Qu. Diagnostic-Guided Dynamic Profile Optimization for LLM-based User Simulators in Sequential Recommendation. AAAI 2026.

Questions for authors
  1. How does the self-review procedure described in the main text correspond to the actual implementation? Equation (6) describes judgments based on a perturbed historical sequence, whereas Appendix F states that only a perturbation plan is provided without actually applying the perturbation.
  2. Does consensual experience adjustment require support from distinct historical events? Each practice sample generates ten proposals, while Appendix E allows a viewpoint supported by three proposals to be incorporated into memory. Must these three proposals come from different practice samples? How are repeated proposals from the same case prevented from being counted as multiple pieces of supporting evidence?
  3. Please clarify the evaluation protocol. How many users and prediction targets are used in each dataset? How are practice samples selected across the 80 rounds?
Reviewer_51xtOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces PraMem, a training-free memory framework for predicting user behavior from long interaction histories. It uses historical actions as practice targets: the LLM makes predictions, compares them with the recorded actions, and proposes changes to its memory. The memory stores both user behavioral patterns and reminders about the model’s potential prediction biases. Two LLM-based checks screen proposals for dependence on historical evidence and specificity to the practice scene. Proposals with repeated support are then incorporated into memory. Experiments on OmniBehavior, MemoryCD, and MovieLens-1M with two prediction backbones show accuracy gains over the evaluated baselines, particularly on OmniBehavior.

Critical strengths

S1. Clear gains on OmniBehavior. PraMem improves accuracy across all four OmniBehavior scenarios with both prediction backbones. These results provide encouraging evidence that feedback from historical prediction tasks can improve memory construction.

S2. A practical approach without weight updates. The method builds user-specific memory offline and reuses it during prediction. Memory constructed with GPT-OSS-120B also improves Qwen’s performance, suggesting that the resulting experience can be useful beyond the model that generated it.

S3. Broad ablation coverage. The ablation studies cover the main components, including the two memory types, reflection, consensus updates, and both self-review checks. These results help assess how each component contributes to overall performance.

Critical weaknesses

W1. Accuracy alone leaves the main results difficult to interpret. OmniBehavior originally evaluates binary behaviors with F1, while MemoryCD evaluates rating prediction with MAE and RMSE. The paper instead reports accuracy throughout, without adequately explaining this change or reporting class distributions and majority-class baselines. It is therefore unclear whether the gains reflect better prediction across behavior classes or mainly better alignment with dominant labels. For rating prediction, accuracy also treats errors of one star and four stars equally. Reporting the original metrics and per-class results would help establish whether the conclusions hold beyond accuracy.

W2. The comparisons do not fully isolate the benefit over a simpler reflection-based memory loop. The paper includes useful ablations that remove individual components, but does not report a jointly simplified prediction–reflection–memory-update baseline without the proposed self-review and consensus mechanisms, using the same practice samples and a comparable construction budget. This comparison would help establish whether the full design offers gains beyond straightforward feedback-driven memory refinement.

W3. The benefit of longer history remains unclear. The paper specifies the prediction input windows and practice schedule, but does not report the full history lengths used for memory construction. Figure 3 varies practice rounds rather than historical coverage, so it does not show whether access to more distant interactions improves prediction. Comparing shorter and longer historical windows with the same prediction targets and number of practice rounds would clarify this point.

Questions for authors
  1. Could the authors explain the choice of accuracy and report F1 for OmniBehavior, MAE/RMSE for rating prediction, and the corresponding class distributions and majority-class baselines? Consistent gains under these metrics would strengthen my confidence in the main results.
  2. Beyond the individual component ablations, could the authors evaluate a jointly simplified prediction–reflection–memory-update baseline using the same historical practice samples and a comparable construction budget? This would help clarify the added value of the proposed self-review and consensus mechanisms.
  3. Does access to longer history improve prediction? Could the authors report the history lengths used for memory construction and compare shorter and longer windows using the same prediction targets and number of practice rounds?
  4. What is the total cost of memory construction? Please report cumulative input/output tokens and LLM calls per user to complement the reported construction time.
Reviewer_rB2mOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PraMem, a training-free framework for long-horizon behavior prediction with LLMs. Instead of managing lengthy historical sequences through context compression, PraMem constructs a user-specific and time-evolving experiential memory, which consists of pattern experience and bias-alert experience, through practice sample construction with segments drawn from the historical sequence, refelective proposal generation and concensual experience adjustment. A self-review mechanism is introduced to filter reflective proposals based on their groundedness in the practice samples and their generalizability. Experiments on OmniBehavior, MemoryCD and MovieLens-1M demonstrate that PraMem outperforms prior methods with GPT-OSS-120B and Qwen3.5-35B-A3B.

Critical strengths
  1. The paper identifies the limitations of memory management through information compression and proposes a reasonable alternative that exploits historical sequences to construct experiential memory through practice.

  2. The proposed self-review mechanism provides an explicit way to assess the quality of reflective proposals based on groundedness and generalizability.

  3. Experiments across multiple benchmarks and two LLM backbones show improvements over existing methods. The contributions of the proposed components are supported by ablation studies.

Critical weaknesses
  1. The groundedness review examines whether a proposal would remain valid after perturbing the historical sequence of a practice sample. However, according to the prompt in the appendix, the perturbation is not actually executed. The LLM is asked to judge whether the proposal would still hold if the perturbation were applied. There may be a gap between the hypothetical effect of a perturbation predicted by the LLM and the outcome of actually applying it. Since this judgment depends on the LLM's ability to reason about hypothetical changes, the reliability of the groundedness review is less convincing.

  2. Sec. 4.1 states that GPT-OSS-120B is used for all PraMem procedures during memory construction. However, it is unclear which LLMs are used to construct the memories for the baselines, such as Mem0, MemOS and ProEx. Although Table 1 evaluates all methods using the same backbones, the choice of LLM for memory construction may also affect performance. Without this information, it is difficult to assess the fairness of the comparison.

  3. The paper focuses on long-horizon behavior prediction, but the experiments do not directly examine how performance changes with the length of the historical sequence. It remains unclear whether comparable improvements could also be obtained with longer or shorter histories.

Questions for authors
  1. Which LLMs are used for memory construction in Mem0, MemOS and ProEx in Table 1? Could the authors also report results where PraMem's experiential memory is constructed using Qwen3.5-35B-A3B rather than GPT-OSS-120B?

  2. How many runs were conducted for PraMem and the baselines in Table 1, and which statistical test was used to determine significance? Could the authors also report the standard deviations across runs, if multiple runs were conducted?

  3. How sensitive is PraMem to the choice of hyperparameters such as the experience adjustment interval TT?

Reviewer_4CwPOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

PraMem builds user-specific textual memory by practicing next-action prediction on historical events with known outcomes. The LLM predicts an action, compares it with the observed outcome, and proposes memory updates. The memory separates user behavioral patterns from reminders about errors the model tends to make. Proposals pass self-review checks and are periodically consolidated through consensus. Final prediction uses recent history plus the resulting memory, without updating model weights.

I find the central idea valuable: the usefulness of customer memory is tested against observed behavior, and the memory is revised based on predictive shortcomings. The paper provides encouraging evidence that this feedback-driven construction procedure improves prediction. I regard the contribution as an empirical memory-management method, with some mechanism claims and evaluation details that should be clarified.

Critical strengths
  1. Historical outcomes provide a useful feedback signal for memory construction. The practice-and-revision loop goes beyond assessing whether a customer profile is readable or plausible. It tests whether the profile helps predict known outcomes and uses mistakes to guide updates. The temporal formulation explicitly places practice samples before the prediction target. This is a practical approach to making customer memory more useful for downstream decisions. (Sections 2 and 4.1.)

  2. The reported improvements are broad and sometimes substantial. PraMem exceeds the best listed baseline in 17 of the 18 backbone/domain settings in Table 1. On OmniBehavior Ads, accuracy improves from 68.1% to 80.8% with GPT-OSS and from 61.3% to 75.4% with Qwen. Gains on MemoryCD are smaller, and GPT-OSS Personal Care is an exception. These settings are correlated rather than independent replications, but the breadth of the positive results still provides useful evidence.

  3. Memory constructed with one model is useful to another. The memory is built with GPT-OSS and reused with Qwen. This supports some portability of the memory representation and is practically relevant when prediction models change without rebuilding every customer's memory. The result applies to the combined memory; it does not by itself establish that every model-specific bias alert transfers. (Sections 4.1-4.2.)

  4. The paper examines the components and reports construction cost. The ablations investigate practice-time reasoning, reflection, consensus, the two memory types, and self-review. The manual audit gives an initial assessment of proposal quality. Table 4 also reports 28.8 minutes per user for construction, making a material practical cost visible. Together, these analyses provide more support than the main accuracy comparison alone. (Figures 4-5 and Tables 2-4.)

Critical weaknesses
  1. Evaluation and audit reporting need more detail. The manuscript should specify evaluated user/target counts, subset selection, label distributions, and the exact historical-prefix/held-out-target protocol. The p < 0.01 markers in Table 1 also need a test description, repeated-run information, and an explanation of how within-user dependence is handled. These details matter especially for small gains. Table 3's manual audit needs its sample size and annotation procedure. The stated temporal safeguards are useful; I did not identify a demonstrated violation, but the reporting should be complete enough to reproduce and assess the results.

  2. The self-review checks should be presented and validated as heuristics. Appendix F describes judging hypothetical perturbation plans rather than instantiated perturbed histories. Appendix G identifies the real scene to the judge, so the procedure is a prompted specificity assessment rather than blinded scene selection. These choices can still produce useful filtering, and the ablations provide encouraging empirical support, but they do not establish guaranteed groundedness or generalization. A valid brand preference could also identify one scene among distractors while generalizing to other products. The main description should match the implementation, and the audit should examine such cases. Consensus should additionally clarify whether supporting proposals must originate from distinct historical events.

  3. The distinctive contribution could be isolated more clearly. A comparison with diagnostic profile refinement such as DGDPO, discussed in Section 5, would strengthen the positioning, with differences in training requirements accounted for. The bias-alert component would also benefit from a control that holds pattern memory fixed and compares the initialized generic reminders with learned user-specific alerts. This would clarify whether gains arise from error-derived guidance or additional user-relevant information. These are limits on the attribution of the improvements; the current evidence for the complete memory-construction procedure remains useful.

Questions for authors
  1. Please provide evaluation population sizes, label distributions, the temporal protocol, significance-test details, and the manual-audit sample and annotation procedure.

  2. How much does learned bias-alert memory add over the same pattern memory with fixed generic reminders? How does the method differ empirically from diagnostic profile refinement, accounting for training requirements?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.