acceptodds
Under review as a conference paper at ICLR 2027

MemoWM: How World Models Change What Agents Need to Remember

Abstract

World models capture regularities that can be reused across experiences, whereas episodic memories incur a new storage cost at every interaction. Making the same world model available at recall allows an agent to reconstruct predictable content and store only the additional information needed to preserve future task quality. We introduce ResidualMem, a task-aware conditional coding framework that turns this shared predictive knowledge into episodic storage savings. The world model provides a coding distribution for retained state components and reconstructs omitted ones. A state-dependent utility gate combines calibrated prediction-error probabilities with the task impact of an error to retain corrections according to their estimated risk and coding cost. Across six long-term agent-memory benchmarks, ResidualMem achieves the highest average QA accuracy and the lowest average persistent storage among the evaluated memory systems. We further find that stronger world models require fewer episodic bits at comparable task quality, but do not always minimize total storage once model parameters are included. The storage-minimizing model size increases with the number of remembered interactions, as savings on each experience offset the additional parameter cost.

Primary area:
foundation or frontier models, including LLMs
Keywords:
agent memory,world model

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_vTf3OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper asks a useful question for long-lived agents: once an agent has a world model that captures regularities of its environment, what information from an individual experience still needs to be stored explicitly? The central idea is that persistent memory should be divided into shared world knowledge and episode-specific information. If the world model can reliably reconstruct some part of an experience, repeatedly storing that information is wasteful; episodic memory should preserve primarily what is both difficult to predict and consequential for future tasks. The proposed framework, MemoWM, implements this idea by encoding observations into a compact, query-independent discrete representation and using an action-conditioned world model to predict each coordinate from previously reconstructed states.

Critical strengths
  • The core idea is both conceptually clean and technically substantive. Treating a world model as shared side information for episodic memory is a natural but nontrivial reframing.
Critical weaknesses

The task-aware componen relies substantially on the offline task/query distribution, and very-long-horizon robustness remains only partially demonstrated.

Questions for authors

N.A.

7 comments
AuthorsOpenReview ↗

Response for Weakness 1

We sincerely thank you for your thoughtful review and thorough understanding of our work. We especially appreciate your recognition of the core idea as conceptually clean and technically substantive, and of the value of treating world models as shared side information for episodic memory. Your insightful feedback on offline task statistics and longer-horizon robustness motivated the held-out analyses and 150-step closed-loop evaluation presented below.



Weakness 1: The task-aware component relies substantially on the offline task/query distribution.



We first clarify the role of the offline statistics and their held-out evaluation, then present the longer-horizon results.

• Why use offline task-impact statistics?

MemoWM is designed to construct query-independent memory, meaning that memory is written before future questions are known. Therefore, task importance cannot be estimated using a specific future query at writing time. Instead, we estimate two shared statistics offline: (1) a coordinate-level task-impact prior Vⱼ, which measures how much a prediction error affects future answers, and (2) an entropy-to-error lookup gⱼ(H), which estimates prediction error probability from predictive entropy. Both are fixed before deployment and shared by the writer and decoder. This allows the same memory to support different future queries without requiring an additional Reader call during memory writing.

These estimates depend on the offline data distribution. We evaluate their usefulness on held-out trajectories and unseen-query variants below. First, the lowest-impact coordinates remain relatively unimportant across different trajectories and query formulations:

Table 3(c). Answer-NLL damage conditioned on prediction errors for original queries and independently constructed unseen-query variants (10⁻³ bits/answer). (Section 4.5 of the revised manuscript.)

Coordinate subsetOriginal queriesUnseen-query variantsRandom subsets
Bottom 5%0.2120.32337.4
Bottom 10%0.2210.37538.8
Bottom 15%0.2310.42338.3
Bottom 20%0.2320.56539.8

Even on independently constructed unseen-query variants, omitting these coordinates causes much less damage than omitting equally sized random subsets.

Second, we evaluate the frozen entropy-to-error lookup on held-out trajectories:

Table 3(b). Frozen entropy lookup versus observed prediction error (%). (Section 4.5 of the revised manuscript.)

Predictive entropy (bits)Frozen lookup (%)Observed error (%)
[0, 0.5)4.08.1
[0.5, 1)9.516.0
[1, 1.5)17.226.2
[1.5, 2)27.135.1
[2, 2.5)36.545.3
[2.5, 3)44.852.5
[3, 4)57.562.5
[4, 5)76.378.8
[5, 6)89.088.8

The observed prediction error increases monotonically from 8.1% to 88.8% as entropy increases, and the frozen lookup preserves the same trend. Although its absolute estimates are not perfectly calibrated, it provides a useful signal for ranking prediction risk on held-out data. Together, these results support reusing offline task-impact and prediction-risk statistics across different states and queries. A substantially different deployment task distribution may require recalibration. We will clarify the conditions for reusing these offline statistics in the revised manuscript.

AuthorsOpenReview ↗

Response for Weakness 2



Weakness 2: very-long-horizon robustness remains only partially demonstrated.



• Robustness over longer interaction horizons.

The original evaluation covers rollouts up to 40 steps. To examine longer horizons, we collected additional WebArena trajectories and extended the closed-loop evaluation to 150 steps, using the same comparison with a teacher-forced oracle:

Table 7. Closed-loop reconstruction relative to a teacher-forced oracle on WebArena rollouts up to 150 steps. Extra NLL is in bits/answer; extra storage is in bits/state. (Appendix G of the revised manuscript.)

Step rangeExtra NLL (bits)Extra storage (bits/state)
1–250.01015
26–500.01419
51–750.01821
76–1000.02223
101–1250.02524
126–1500.02725

Both additional NLL and storage increase gradually over the 150-step rollout, without a sharp deterioration. Even in the final 25 steps, the additional NLL is only 0.027 bits, with 25 additional bits/state of storage. These results extend our original 40-step evaluation (revised manuscript, Section 4.6; Figure 4(b)) to 150 steps and provide further evidence of stable closed-loop reconstruction over longer interactions. Thank you again for highlighting these two aspects of robustness. Your feedback helps us clarify when the offline statistics remain useful and the interaction horizons covered by our evaluation.

We hope these clarifications and additional results address your concerns, and we look forward to your further feedback.

Dear Reviewer vTf3,

Thank you again for your careful review and constructive suggestions. We would like to briefly follow up on our responses and the additional experiments.

When you have a chance, we would be grateful to hear whether the new results and clarifications address your main concerns, or whether any points would benefit from further explanation. We would also appreciate an updated assessment of the paper in light of these additions and revisions.

Thank you for your time and consideration.

Best regards, The Authors

Reviewer_f8MROpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

this works propose an interesting idea -- using a world model (WM) to help memory management. WM predicts regularities (expected things), memory stores exceptions. in this way, the "normal/expected" information can be omitted/compressed.

Critical strengths

the way this work utilizes WM seems novel to me. they present a nice integration of WM and memory management, draw a meaningful boundary between different ways to store information of different nature.

Critical weaknesses

How much of MemoWM's improvement comes from actually learning what actions do, versus simply learning that most of the environment stays the same? As for many steps in the environment being tested, the majority elements stay the same after an action is taken, probably a baseline memory management is just to store the difference?

Questions for authors

as mentioned above, is a learned world model really necessary? my intuition says yes, but probably we need more empirical support.

7 comments
AuthorsOpenReview ↗

Response for Weakness

We sincerely thank you for your thoughtful review and positive assessment of our work. We especially appreciate your recognition of the novel integration of world models and memory management, and of the meaningful boundary between predictable information and experience-specific content. Following your insightful suggestion, we added difference-based baselines to test whether the gains extend beyond temporal persistence, and report them alongside controlled ablations below.



Weakness: How much of MemoWM's improvement comes from actually learning what actions do, versus simply learning that most of the environment stays the same? As for many steps in the environment being tested, the majority elements stay the same after an action is taken, probably a baseline memory management is just to store the difference?



Temporal persistence makes storing differences between consecutive states a natural baseline. We consider two baselines: Dense-mask Delta stores a 1,024-bit change mask per transition, while Sparse-index Delta stores a 10-bit index per changed coordinate; both store the new values of all changed coordinates. Following your suggestion, we conducted additional experiments on two benchmarks.

Table 4. Difference-based memory versus MemoWM. Rate is state-code storage in bits/state, including initial states and delta masks or indices; overhead is relative to MemoWM within each benchmark. Accuracy is in %. (Appendix C.4 of the revised manuscript.)

BenchmarkChange rateMethodState-code rate (bits/state)Accuracy (%)
WorldMemArena87%MemoWM4,33560.18
Dense-mask Delta6,347 (+46.4%)60.27
Sparse-index Delta13,103 (+202.3%)60.27
LongMemEval-V246%MemoWM1,24442.91
Dense-mask Delta4,184 (+236.3%)42.98
Sparse-index Delta7,334 (+489.5%)42.98

The tested difference-based baselines have substantially higher state-code rates on both benchmarks, while achieving only marginally higher accuracy (0.09 and 0.07 percentage points, respectively). On LongMemEval-V2, MemoWM uses 1,244 bits/state, compared with 4,184 for Dense-mask Delta and 7,334 for Sparse-index Delta, corresponding to 236.3% and 489.5% overhead, respectively. Notably, although LongMemEval-V2 has a lower change rate (46%), the relative state-code storage advantage of MemoWM is even larger. This suggests that MemoWM's benefits are not limited to environments with frequent state changes.

• Why does MemoWM achieve better storage efficiency?

A straightforward difference-based memory needs to tell the decoder which coordinates have changed at every step. This requires storing an additional binary mask or the indices of changed coordinates, besides the new values themselves. For our 1,024-dimensional state representation, a dense binary mask alone requires 1,024 bits per transition. Although sparse indexing can reduce this cost, it still requires 10 bits for each changed coordinate, since there are 1,024 possible coordinate indices (log₂(1024) = 10). In contrast, MemoWM does not need to store or transmit a retention mask for each observation. Given the same reconstructed history, action, and shared world model, the writer and decoder can independently compute the predictive distribution and retention decisions at each step. Therefore, the decoder knows which coordinates to retrieve and which to reconstruct without receiving an additional mask. The shared-mask construction is described in Section 3.2 of the revised manuscript.

The key difference is that Delta Memory only distinguishes between changed and unchanged coordinates, and must store every changed value. It does not consider whether a changed value can already be predicted from previous states and actions, or whether reconstructing it incorrectly would affect future tasks. MemoWM considers both. It uses the shared world model to predict each coordinate and estimates the impact of prediction errors on future tasks. A coordinate can therefore be omitted even if its value has changed, as long as it can be reliably predicted and the expected task impact of omitting it is low. This allows MemoWM to allocate storage based on both predictability and task importance, rather than simply whether a coordinate has changed.


AuthorsOpenReview ↗

Response for Question



Question As mentioned above, is a learned world model really necessary? My intuition says yes, but probably we need more empirical support.



Thank you for asking us to examine the contribution of the learned world model more directly. To isolate the contribution of learned dynamics beyond temporal persistence, we examine action/history ablations and compare world-model prediction with previous-state copying. The additional difference-based baselines are reported in our response to Q1 and in Appendix C.4, Table 4 of the revised manuscript.

• Contribution of action and history information.

We compare the full predictive context with variants that remove action or history information, while retaining all coordinates:

Figure 3(c). All-coordinate rate under four predictive contexts. (Selected contexts, tabulated from Section 4.4 of the revised manuscript; rates are in bits/transition.)

World modelFull contextNo actionNo history
Small3,8314,064 (+6.1%)3,960 (+3.4%)
Large3,3863,857 (+13.9%)4,249 (+25.5%)

These ablations show that the world model reduces coding cost by exploiting action-conditioned dynamics and historical context. Removing action or history increases the coding rate by 6.1% and 3.4% for the small model, respectively, versus 13.9% and 25.5% for the large model. The larger penalties for the large model indicate that greater world-model capacity enables more effective use of the same actions and history for predictive coding.


• World-model prediction versus previous-state copying and static-prior fill.

We compare world-model prediction (WM) with two alternatives for reconstructing omitted coordinates. Previous-state copying (Prev) copies the corresponding values from the previous state. Static-prior fill (Prior) fills the omitted coordinates using a fixed prior, independent of the current action and history:

Figure 4(a). Reconstruction quality. (Selected methods and metrics, tabulated from Section 4.6 of the revised manuscript.)

MetricPrevious-state copying (Prev)Static-prior fill (Prior)World-model prediction (WM)Improvement vs. Prev
Answer-NLL distortion (bits) ↓0.12800.13010.086732.3% lower
Semantic match (%) ↑93.892.397.0+3.2 percentage points
Answer accuracy (%) ↑57.1055.6060.18+3.08 percentage points

With the omitted coordinates and storage budget held fixed, WM outperforms both Prev and Prior on all three metrics. In particular, answer accuracy reaches 60.18%, compared with 57.10% for Prev and 55.60% for Prior. These gains show that predicting omitted information from actions and history preserves more evidence for subsequent answers than copying previous values or using a static prior.

We hope these clarifications and additional results address your concerns, and we look forward to your further feedback.

Dear Reviewer f8MR,

Thank you again for your careful review and constructive suggestions. We would like to briefly follow up on our responses and the additional experiments.

When you have a chance, we would be grateful to hear whether the new results and clarifications address your main concerns, or whether any points would benefit from further explanation. We would also appreciate an updated assessment of the paper in light of these additions and revisions.

Thank you for your time and consideration.

Best regards, The Authors

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

This paper investigates whether world models can reduce the amount of episodic information that long-term agents need to store. The central idea is that information predictable from a shared world model need not be stored explicitly, provided that its reconstruction errors do not significantly affect downstream tasks.

The authors propose MemoWM, a memory framework combining quantized state representations, predictive reconstruction, task-aware information retention, and shared-history reconstruction. Its retention mechanism considers both reconstruction uncertainty and the potential impact of errors on downstream answers. The framework also examines how world-model capacity influences the trade-off between shared model parameters and experience-specific storage.

Critical strengths
  1. Broad empirical evaluation. The comparison covers five existing benchmarks, twelve baselines, and multiple memory paradigms. Table 1 reports favorable aggregate accuracy and storage efficiency, providing evidence that the proposed approach is competitive.

  2. Useful ablation and scaling analyses. Table 2 distinguishes the contributions of quantization, predictive coding, and task-aware omission. Figure 3 also provides an interesting analysis of how larger world models can reduce recurring episodic storage and how that benefit changes with the number of interactions.

Critical weaknesses
  1. The practical efficiency gains are not fully established. The main storage comparison excludes shared model weights, while MemoWM requires additional components for representation learning, predictive reconstruction, and retention decisions. Although Figure 3 considers amortized model storage, the paper does not provide a comparable end-to-end analysis of training cost, inference latency, memory-writing overhead. These omissions make it difficult to establish whether the proposed storage savings translate into lower overall operational cost.

  2. The advantages of task-aware retention over simpler alternatives are relatively modest. Figure 4 reports a comparatively small separation between the proposed factorized risk estimator and simpler gating strategies. This raises questions about whether the additional complexity of the risk-estimation mechanism is justified by its marginal benefits.

  3. The experimental evidence does not fully establish generality across world models and agent settings. The principal benchmarks assess memory-dependent question answering rather than the success of agents performing extended interactive tasks. Moreover, the model-capacity experiments in Figure 3 are conducted on AMA-Bench using a specialized predictive model. The observed relationship between world-model capacity and storage requirements may not transfer directly to pretrained foundation models, unfamiliar environments, or changing environment dynamics.

Questions for authors
  1. Retention mechanism: How substantial and statistically reliable are the improvements of factorized risk estimation over simpler gating alternatives? Would the additional complexity remain justified under a fixed overall computation budget?
8 comments
AuthorsOpenReview ↗

Response for Weakness 1

We sincerely thank you for your careful reading and constructive feedback. We appreciate your positive assessment of the empirical evaluation and ablation studies, as well as your suggestions for strengthening the efficiency analysis and clarifying the scope of our conclusions. Below, we respond to each concern in turn.



Weakness 1: The practical efficiency gains are not fully established. The main storage comparison excludes shared model weights, while MemoWM requires additional components for representation learning, predictive reconstruction, and retention decisions. The paper also does not provide a comparable end-to-end analysis of training cost, inference latency, or memory-writing overhead.



Thank you for highlighting the need to distinguish storage savings from the costs of the additional model components. Following your suggestion, we have additionally compared memory writing and reading latency with five baselines. Below, we clarify the storage accounting, present the runtime comparison, and report the world-model training and computation costs.

• Storage accounting.

Our main storage comparison measures the persistent memory payload required to retain each experience, rather than the parameters of the models used to construct or access that memory. As defined in Appendix C.1 of the revised manuscript, functional model weights for encoding, embedding, extraction, retrieval, and reading are excluded from per-experience storage. This convention is applied consistently across all methods. For example, the embedding model used by RAG and the extraction model used by A-Mem are also excluded. This does not mean that these model components are cost-free; rather, they are shared functional parameters instead of experience-specific memory content. We acknowledge that this distinction should have been made clearer in the main paper.

For completeness, the additional shared components in MemoWM, including the Q-Former and retrieval head but excluding the world model, have a combined parameter footprint of 109.23 MiB (revised manuscript, Appendix F, Table 6). The world-model parameters are accounted for separately in Section 4.4 and Figure 3(b) of the revised manuscript, where their FP16 storage cost is amortized over retained experiences. This accounting convention and the additional parameter footprint are now documented in Appendices C.1 and F of the revised manuscript.


• Measured memory writing and reading latency.

MemoWM achieves the lowest combined writing and reading latency (0.206 s) among the nine methods evaluated with the same framework and timing boundaries. Write covers memory construction and persistence; read covers query encoding, memory search, and entry retrieval. Answer generation by Qwen3.5-9B is excluded for all methods.

Table 5. Measured memory-operation latency. Answer generation is excluded for every method. (Appendix F of the revised manuscript.)

MethodTotal (s)Write (s/session)Read (s/query)
MemoWM (ours)0.2060.1390.067
M2A0.2940.2860.008
RAG0.3010.2910.010
UniversalRAG0.7600.2950.465
MemGPT1.164~0.00011.164
AUGUSTUS3.4401.6131.827
MemAgent24.80424.8000.004
A-Mem64.59364.50.093
MIRIX133.935133.8000.135

• World-model training and forward computation.

The training cost covers only the world model; computation is reported per forward pass. Memory writing and decoder-side reconstruction each require one pass per observation.

Table 6. Component-level fitting, computation, and parameter-storage costs. Training costs are one-time offline costs; the WM forward cost is for one pass in either writing or reconstruction. MiB denotes 2²⁰ bytes. (Selected entries from Appendix F of the revised manuscript.)

MetricCost
World model training200.0 GPU-hours
World model forward computation31.2 GFLOPs/observation

• Physical storage validation.

On WorldMemArena Web, the actual entropy-coded stream uses 4,388.70 bits/state, only 1.50% above the ideal estimate of 4,323.84 bits/state. Independent decoding exactly reproduces 100% of the writer-side reconstructed states (Appendix D of the revised manuscript).

Thank you for suggesting these additional efficiency measurements.


AuthorsOpenReview ↗

Response for Weakness 2 & Question



Weakness 2 & Question: The advantages of task-aware retention over simpler alternatives are relatively modest. Figure 4 reports a comparatively small separation between the proposed factorized risk estimator and simpler gating strategies. This raises questions about whether the additional complexity of the risk-estimation mechanism is justified by its marginal benefits.



Thank you for asking whether the benefits of task-aware retention justify its additional complexity. To assess this trade-off, we compare MemoWM with simplified retention rules and measure the additional computational cost.

• How does the factorized risk estimator work?

MemoWM estimates the omission risk of each coordinate by combining two factors: the probability of a prediction error (eₜ,ⱼ), estimated from predictive entropy, and its downstream task impact (Vⱼ), estimated from offline data. A coordinate is retained when its estimated omission risk exceeds its weighted coding cost. Since both the writer and decoder share the same reconstructed history, world model, and offline statistics, they independently reproduce the same retention decisions without transmitting a per-state mask. The shared-mask construction is described in Section 3.2 of the revised manuscript.


• Comparison with simplified retention rules.

We compare the full estimator with four simplified retention rules at the same closed-loop reconstruction distortion (|ΔNLL| = 1.536). Static-e and Uniform-V are component ablations that remove the state dependence of prediction risk and the coordinate specificity of task impact, respectively. Uncertainty uses predictive entropy alone, while Random uses random gating.

Table 3(a). Gating at matched distortion; rate is bits/state, and Rate ↑ is the increase over MemoWM. (Section 4.5 of the revised manuscript.)

Retention ruleStorage (bits/state)Overhead
MemoWM2,354.2–
Static-e2,754.3+17.0%
Uniform-V2,954.8+25.5%
Random3,134.2+33.1%
Uncertainty3,342.1+42.0%

MemoWM achieves the lowest storage at matched distortion. In particular, removing either component increases storage by 17.0% or 25.5%, respectively. This shows that both state-dependent prediction risk and coordinate-specific task impact contribute to more efficient retention.


• Why are both factors important?

The factorization is motivated by the decomposition of expected omission loss into the probability of a prediction error and its expected task impact given an error (Appendix A.5 of the revised manuscript). Our held-out experiments in Section 4.5 of the revised manuscript (Table 3(b,c)) further support both factors:

  • Prediction risk: The observed prediction-error rate increases monotonically from 8.1% to 88.8% across entropy bins (Table 3(b) of the revised manuscript), supporting predictive entropy as a useful signal for estimating state-dependent risk.
  • Task impact: Omitting the lowest-impact 5%–20% of coordinates causes only 0.212–0.232 × 10⁻³ bits/answer of error-conditioned damage, compared with 37.4–39.8 × 10⁻³ for equally sized random subsets (166–176× higher; Table 3(c) of the revised manuscript).

These results support the complementary roles of the two factors: one estimates how likely a prediction error is, while the other estimates how much that error matters.


• Computational overhead.

We also measured the additional cost of the retention mechanism:

Table 6. Component-level fitting, computation, and parameter-storage costs. Training costs are one-time offline costs; the WM forward cost is for one pass in either writing or reconstruction. MiB denotes 2²⁰ bytes. (Selected entries from Appendix F of the revised manuscript.)

MetricCost
Additional online gating latency0.2 ms/observation
Offline risk estimation and calibration20 GPU-hours

The online overhead is negligible, as gating requires only entropy-based lookups and lightweight coordinate-wise operations, without an additional world-model forward pass or Reader call. The 20 GPU-hours are a one-time offline cost rather than a recurring deployment cost. Overall, although the gains over individual simplified rules are moderate, both components consistently improve storage efficiency at matched reconstruction quality, with negligible online overhead. We believe this is a favorable trade-off for long-lived agents, where the storage savings accumulate over time. We appreciate this suggestion, which helps us present the benefits and costs of the retention mechanism more concretely.

AuthorsOpenReview ↗

Response for Weakness 3



Weakness 3: The experimental evidence does not fully establish generality across world models and agent settings. The main benchmarks assess memory-dependent question answering rather than the success of agents performing extended interactive tasks. Moreover, the model-capacity experiments in Figure 3 are conducted on AMA-Bench using a specialized predictive model, so the observed relationship may not transfer directly to pretrained foundation models, unfamiliar environments, or changing environment dynamics.



Thank you for highlighting the distinction between memory-dependent QA and interactive agent success, and for asking how our findings extend across world models and environments. Below, we present additional evaluations and clarify the scope of our findings.

• Agent-level evaluation.

Memory-dependent QA does not directly measure agent task success. MemoWM can also be integrated into an online agent loop: after each interaction step, the completed observation is written to memory, and its reconstructed representation is available for subsequent decisions. The agent continues to receive the actual current observation, while its past experience is retrieved from reconstructed memory.

To evaluate whether MemoWM supports actual agent behavior, we additionally compare MemoWM with A-Mem, MemRL, and ACE on WebShop:

MethodTask success (%)Memory storage (bits/observation)
A-Mem15.4114,849
MemRL14.03,268
ACE16.012,031
MemoWM (ours)17.82,687

MemoWM achieves the highest task success among the compared methods. It also uses the least memory storage, a 17.8% reduction relative to MemRL. These results extend our evaluation beyond memory-dependent QA to interactive agent task success while showing lower storage requirements.


• Capacity scaling under cross-benchmark evaluation.

In our original capacity-scaling study (Figure 3 of the revised manuscript), world models of different capacities are trained and evaluated using AMA-Bench data. To examine whether the observed capacity–storage relationship depends on this particular training distribution, we train world models of different sizes on WorldMemArena and evaluate them on held-out AMA-Bench trajectories. We measure both next-state prediction quality and the corresponding coding rate:

WM parametersPrediction accuracy (%)Coding rate (bits/transition)
Small15.75,421
Medium18.55,132
Large22.14,950

Across the tested capacities, prediction accuracy increases from 15.7% to 22.1%, while the coding rate decreases from 5,421 to 4,950 bits/transition on AMA-Bench, despite the world models being trained on WorldMemArena. These results provide additional evidence that the capacity–storage relationship is not specific to training on AMA-Bench, although the magnitude of the benefit depends on the training and evaluation distributions.


• Cross-benchmark transfer.

We further evaluate the transferability of predictive coding by training world models on different benchmarks and measuring their coding rates on AMA-Bench and WebQuest. To isolate the contribution of the predictive prior, task-utility gating is disabled in this experiment.

Table 8. Cross-benchmark WM coding rates with task-utility gating disabled (bits/transition). Parentheses indicate increases over the in-domain model in the same evaluation column. (Appendix H of the revised manuscript.)

WM training data \ Evaluation dataAMA-Bench (evaluation)WebQuest (evaluation)
AMA-Bench4,1233,568 (+52.9%)
WorldMemArena4,953 (+20.1%)3,124 (+33.8%)
WebQuest4,879 (+18.3%)2,334

The results show that predictive coding transfers across the tested benchmarks, although cross-benchmark models generally incur higher coding rates than in-domain models. These results suggest that the predictive prior captures some reusable transition structure beyond its training benchmark, while also showing that its effectiveness depends on the target data distribution.


• Clarifications regarding pretrained world models and changing dynamics.

Our experiments use task-trained, action-conditioned predictors designed for MemoWM's state representation. Extending MemoWM to pretrained foundation world models requires adapting their representations and prediction interfaces to support synchronized prediction and reconstruction at both the writer and decoder. Furthermore, although our cross-benchmark experiments demonstrate transfer across the tested data distributions, they do not establish robustness to arbitrary unseen environments or dynamically changing transition rules. We will clarify the evaluated model settings and the scope of the cross-benchmark evidence.

We hope these clarifications and additional results address your concerns, and we look forward to your further feedback.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Reviews, Revisions, and Additional Experiments

Dear PCs, SACs, ACs, and Reviewers,

We thank you and the reviewers for your time and constructive feedback. Through additional experiments, revisions, and clarifications, we have addressed most of the concerns raised by the reviewers.

The reviewers recognized the novel integration of world models and memory management (f8MR), the “conceptually clean and technically substantive” framing (vTf3), and the evaluation across five benchmarks and twelve baselines, with ablation and scaling analyses (PRqd).

MemoWM treats a world model as shared knowledge for episodic memory, retaining information that is difficult to reconstruct and useful for future tasks. We find that stronger world models reduce storage per experience, while the model size that minimizes total storage increases as the agent accumulates more experience.

1. Benefits beyond temporal persistence (f8MR). New comparisons on WorldMemArena and LongMemEval-V2 show that Dense-mask Delta requires 46.4% and 236.3% more state-code storage than MemoWM, respectively, with accuracy differences below 0.1 percentage points. At fixed storage and omitted coordinates, world-model reconstruction also reduces answer-NLL distortion by 32.3% over previous-state copying (Appendix C.4, Table 4; Figure 4(a) of the revised manuscript).

2. Task-aware retention and its cost (PRqd). At matched distortion, simplified gating rules require 17.0%–42.0% more storage than MemoWM. The additional online gating cost is 0.2 ms/observation, without another world-model forward pass or Reader call (Section 4.5, Table 3(a); Appendix F, Table 6 of the revised manuscript).

3. Practical efficiency and storage accounting (PRqd). In a comparison with eight baselines under common timing boundaries, MemoWM has the lowest reported combined writing and reading latency (0.206 s; answer generation excluded). Shared parameters, training, and forward computation are reported separately. Physical entropy coding adds only 1.50% over the ideal estimate and exactly preserves writer-side reconstructed states (Appendices C.1, D, and F of the revised manuscript).

4. Interactive agents and cross-benchmark evaluation (PRqd). On WebShop, MemoWM achieves the highest task success (17.8%, versus 16.0% for ACE) and the lowest storage (2,687 bits/observation). Additional capacity-scaling experiments show that larger world models trained on WorldMemArena improve prediction and reduce coding rates on held-out AMA-Bench trajectories. We also report predictive-prior transfer across benchmarks and its additional coding cost (response to PRqd's Weakness 3; Appendix H, Table 8 of the revised manuscript).

5. Offline statistics and longer horizons (vTf3). Held-out trajectories and unseen-query variants support reuse of the offline task-impact prior and prediction-risk ranking on the tested distributions. We also extended closed-loop evaluation from 40 to 150 steps. In the final 25-step range, the additional cost relative to a teacher-forced oracle is 0.027 bits/answer of NLL distortion and 25 bits/state of storage (Section 4.5, Table 3(b,c); Appendix G, Table 7 of the revised manuscript).

Once again, we would like to express our sincere gratitude to the reviewers, ACs, SACs, and PCs for their careful reading and constructive feedback. Thank you very much for your time and consideration.

Sincerely,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.