acceptodds
Under review as a conference paper at ICLR 2027

MemTime: Hippocampus-Inspired Agents for Adaptive Time Series Forecasting with Evolving Memory

Abstract

Pre-trained time-series foundation models have recently achieved remarkable zero-shot forecasting performance across diverse domains, providing a powerful paradigm for time-series prediction. However, most TSFMs remain fixed after pretraining, limiting their ability to incorporate new experiences as data distributions evolve. Inspired by hippocampal memory mechanisms, which enable rapid experience encoding, selective recall, and long-term consolidation, we investigate how dynamic memory processes can endow time-series foundation models with continual adaptation capabilities. Realizing such adaptation requires a mechanism that can effectively identify, incorporate, and evolve experiences throughout streaming forecasting. Motivated by these observations, we introduce MemTime, an autonomous self-evolving time-series agent framework that enables foundation models to adapt through dynamic memory augmentation without modifying their underlying parameters. Specifically, MemTime introduces a self-evolving memory mechanism that enables foundation models to acquire, utilize, and consolidate deployment-time experiences. It incorporates a reinforcement learning-based Cognitive Sentinel, which captures environmental changes and adaptively determines when memory retrieval is required. The Mental Retracer transforms retrieved experiences into memory-based correction signals, which adjust the forecasting process without updating the model parameters. To support long-term adaptation, a memory metabolism mechanism periodically consolidates the memory buffer by removing outdated experiences while preserving valuable knowledge. Extensive experiments on multiple non-stationary forecasting benchmarks demonstrate that MemTime consistently outperforms state-of-the-art adaptation and external-memory baselines, achieving improvements of up to 26.03% while maintaining cross-domain generalization. The code is publicly available at https://anonymous.4open.science/r/MemTime-050A/.

Primary area:
learning on time series and dynamical systems
Keywords:
Adaptive Time Series Forecasting, Hippocampus-Inspired Agents, Evolving Memory

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_SAMUOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MemTime, a hippocampus-inspired framework for adaptive time-series forecasting under evolving data distributions. Built upon a frozen pre-trained time-series foundation model, MemTime introduces three core components: the Cognitive Sentinel, which determines when memory retrieval is needed; the Mental Retracer, which retrieves historical experiences to generate residual correction signals for refining forecasting trajectories; and the Memory Metabolism mechanism, which continuously maintains and updates the memory pool. Together, these components aim to improve forecasting adaptability without updating the backbone parameters.

Critical strengths
  1. The paper addresses a practically meaningful problem of adapting pre-trained time-series foundation models to evolving data distributions. Improving forecasting adaptability without updating the backbone parameters is particularly relevant to real-world streaming forecasting scenarios.

  2. The proposed framework integrates adaptive memory invocation, memory-based forecasting trajectory refinement, and dynamic memory maintenance into a unified forecasting pipeline, demonstrating a clear overall architecture and a well-reasoned methodological design.

Critical weaknesses

1. The main experiments compare MemTime with several time-series foundation models evaluated under zero-shot settings. However, MemTime additionally uses target-domain validation data to train its auxiliary modules, including the residual experts and memory retrieval network. This introduces additional target-domain supervision that is not available to the compared foundation models. Consequently, it remains unclear whether the observed performance improvements primarily stem from the proposed evolving memory mechanism or the additional supervised adaptation.

2. Although the paper cites related studies such as MemCast, TMoE, and STanHop, it does not sufficiently clarify the substantive differences between their core mechanisms and those of MemTime, nor does it provide adequate experimental comparisons with the most closely related methods.

3. The comparisons with recent time-series forecasting methods are insufficient. The main experiments primarily focus on a limited set of pre-trained time-series foundation models, without including several representative recent forecasting methods. Consequently, the experimental results do not provide sufficiently convincing evidence of the proposed method's advantages over recent state-of-the-art approaches.

Questions for authors

Q1. MemTime primarily consists of three components: Cognitive Sentinel, Mental Retracer, and Memory Metabolism. However, these components are not particularly novel. I believe the main contribution of this work lies in integrating them into a coherent and theoretically well-motivated framework. Nevertheless, the ablation results raise a question about whether the Memory Metabolism component is truly necessary.

Q2. Does the retrieval-based combination of LoRA residual experts offer clear advantages over conventional Mixture of Experts (MoE), static LoRA adaptation, or residual prediction networks?

Q3. In the main experiments, MemTime is compared with several time-series foundation models evaluated under zero-shot settings. However, MemTime additionally uses target-domain validation data to train its auxiliary modules.

Q4. According to Appendix A.3, the authors use the target-domain validation set to train the residual experts and memory retrieval network of the Mental Retracer. However, in conventional experimental protocols, the validation set is typically used for model selection and hyperparameter tuning rather than supervised parameter optimization. Therefore, the validation set effectively serves as an additional training set in the current experimental setup. Could the authors further explain the necessity and rationale behind this data partitioning strategy?

Furthermore, the residual supervision signals for the Mental Retracer appear to be directly constructible from historical input windows and their corresponding ground-truth forecasting targets in the training set. Could the authors further clarify why the validation set is used for residual clustering and expert training instead of the training set?

Note: Q3 and Q4 are my primary concerns regarding the overall evaluation and rating of this paper, and the authors' responses to these questions will be key factors in determining whether I reconsider my score.

Specifically, in the main experiments, MemTime additionally uses target-domain validation data to train its auxiliary modules, whereas several baseline models are evaluated under zero-shot settings. Consequently, the current experimental results do not sufficiently distinguish whether the observed performance improvements stem from the additional target-domain supervision or from the proposed evolving memory mechanism.

If I have misunderstood the experimental protocol, I would appreciate further clarification from the authors regarding the rationale behind the current comparisons. More importantly, I encourage the authors to provide additional experimental evidence.

9 comments
AuthorsOpenReview ↗

Response to Reviewer SAMU [PART 1/3]

Thank you very much for the invaluable and in-depth comments. We have tried our best to address your problems by providing more experiments and clarifications as follows.

W1 & Q1. Novelty and Necessity of the Proposed Memory Framework

To the best of our knowledge, MemTime is the first to integrate hippocampus-inspired memory mechanisms with residual velocity correction for continual adaptation of flow-matching time-series foundation models. It represents historical experiences as residual experts and directly incorporates them into the frozen forecasting dynamics.

The proposed framework integrates three complementary mechanisms to ensure that historical experiences are properly invoked, effectively utilized, and continuously maintained. Specifically, the Cognitive Sentinel determines when historical experiences should be invoked, the Mental Retracer transforms retrieved experiences into residual corrections to refine forecasting trajectories, and Memory Metabolism evaluates memory utility to preserve valuable experiences and remove outdated ones. Together, these components enable effective and continual memory-based forecasting adaptation.

More details about the data split have been clarified and explained in our responses to follow-up questions.

W2. Differences across MemTime, MemCast, TMoE, and STanHop.

The differences across MemTime, MemCast, TMoE, and STanHop lie in methodology and performance perspectives, which can be seen as follows.

  • Methodological Differences: MemCast uses historical experiences to guide LLM-based reasoning, TMoE dynamically adjusts experts under distribution shifts, and STanHop retrieves temporal representations through sparse associative memory. In contrast, MemTime represents historical experiences as residual experts and directly integrates retrieved corrections into the velocity field of frozen flow-matching models.
  • Performance Differences: We provide additional experimental comparisons. Specifically, we adapt representative mechanisms from these methods to the same forecasting setting with consistent data partitions and evaluation protocols. MemTime achieves the lowest MSE on six of the seven datasets at the 720-step prediction horizon. On Exchange, it reduces average MSE by 17.5% compared with the baseline TMoE, supporting the effectiveness of retrieving historical residual experts. Full results are provided in Table 1 of Part 2.
AuthorsOpenReview ↗

Response to Reviewer SAMU [PART 2/3]

Table 1: Performance Comparison with TMoE, MemCast, and STanHop

DatasetPredMemTime (MSE/MAE)Sundial (MSE/MAE)TMoE (MSE/MAE)MemCast (MSE/MAE)STanHop (MSE/MAE)
ETTm1960.293 / 0.3400.292 / 0.3420.313 / 0.3530.354 / 0.3750.296 / 0.362
1920.328 / 0.3710.337 / 0.3760.362 / 0.3820.392 / 0.3930.355 / 0.395
3360.368 / 0.4040.370 / 0.4070.391 / 0.4120.425 / 0.4140.385 / 0.427
7200.407 / 0.4250.418 / 0.4330.452 / 0.4430.486 / 0.4480.453 / 0.460
Avg0.349 / 0.3850.354 / 0.3900.380 / 0.3980.414 / 0.4080.372 / 0.411
ETTm2960.174 / 0.2600.178 / 0.2600.179 / 0.2620.183 / 0.2660.177 / 0.266
1920.229 / 0.2910.235 / 0.3040.232 / 0.2970.247 / 0.3050.243 / 0.308
3360.270 / 0.3320.287 / 0.3420.293 / 0.3340.307 / 0.3420.302 / 0.346
7200.336 / 0.3700.360 / 0.3900.394 / 0.3940.407 / 0.3980.402 / 0.404
Avg0.252 / 0.3130.265 / 0.3240.275 / 0.3220.286 / 0.3280.281 / 0.331
ETTh1960.335 / 0.3750.341 / 0.3810.373 / 0.3940.385 / 0.3930.340 / 0.387
1920.379 / 0.4010.381 / 0.4080.422 / 0.4220.435 / 0.4220.398 / 0.437
3360.408 / 0.4170.405 / 0.4240.465 / 0.4480.475 / 0.4440.450 / 0.472
7200.424 / 0.4500.433 / 0.4580.477 / 0.4650.463 / 0.4590.512 / 0.511
Avg0.387 / 0.4110.390 / 0.4180.434 / 0.4320.440 / 0.4300.425 / 0.452
ETTh2960.267 / 0.3230.272 / 0.3320.285 / 0.3360.290 / 0.3390.300 / 0.353
1920.327 / 0.3660.329 / 0.3740.362 / 0.3910.377 / 0.3910.391 / 0.408
3360.354 / 0.3920.357 / 0.3990.415 / 0.4310.416 / 0.4250.435 / 0.448
7200.386 / 0.4480.401 / 0.4420.426 / 0.4490.417 / 0.4460.440 / 0.462
Avg0.334 / 0.3820.340 / 0.3870.372 / 0.4020.375 / 0.4000.392 / 0.418
ECL960.148 / 0.2300.134 / 0.2310.145 / 0.2450.199 / 0.2760.157 / 0.259
1920.164 / 0.2630.154 / 0.2510.163 / 0.2620.198 / 0.2770.227 / 0.304
3360.174 / 0.2710.182 / 0.2760.174 / 0.2720.213 / 0.2930.317 / 0.369
7200.215 / 0.3070.220 / 0.3110.217 / 0.3110.254 / 0.3250.435 / 0.447
Avg0.175 / 0.2680.173 / 0.2670.175 / 0.2730.216 / 0.2930.284 / 0.345
Weather960.160 / 0.2230.158 / 0.2060.153 / 0.2280.165 / 0.2120.263 / 0.302
1920.207 / 0.2510.205 / 0.2530.201 / 0.2550.217 / 0.2960.332 / 0.393
3360.252 / 0.2980.254 / 0.2990.261 / 0.3580.276 / 0.3360.369 / 0.475
7200.315 / 0.3360.322 / 0.3400.347 / 0.3450.339 / 0.3800.456 / 0.526
Avg0.234 / 0.2770.235 / 0.2750.241 / 0.2970.249 / 0.3060.355 / 0.424
Exchange960.147 / 0.2050.237 / 0.2690.148 / 0.2160.183 / 0.2260.192 / 0.211
1920.230 / 0.3750.362 / 0.4150.246 / 0.3950.277 / 0.2960.295 / 0.312
3360.422 / 0.5090.558 / 0.5980.393 / 0.4570.653 / 0.5230.446 / 0.524
7200.562 / 0.6520.683 / 0.7820.861 / 0.6901.014 / 0.7530.862 / 0.765
Avg0.340 / 0.4350.460 / 0.5160.412 / 0.4400.532 / 0.4500.449 / 0.453
AuthorsOpenReview ↗

Response to Reviewer SAMU [PART 3/3]

W3. Comparisons with Recent State-of-the-art Forecasting Methods

We extend the experimental comparisons to include recent time-series methods, including COSA[1], TAFAS[2], and ELF[3] for test-time and online adaptation, as well as TEDM[4] and SARAF[5] for diffusion-based and retrieval-augmented forecasting. The full results are shown in Table 1 of "Response to Reviewer n4uz [PART 1/2]". MemTime achieves the lowest MSE and MAE across all six datasets at both prediction lengths of 336 and 720. At 720 steps, it reduces MSE by 11.2% on ECL compared with the strongest baseline, TEDM, supporting our design for long-horizon forecasting.

Q2. Advantages of Retrieval-based LoRA Residual Experts over Static LoRA, Conventional MoE, and Residual Prediction Networks

We clarify the design motivation and experimental evidence as follows. The method comparison and forecasting results are presented in Tables 1 and 2, respectively, of "Response to Reviewer jieF [PART 2/3]".

  • Retrieval-based Expert Selection. As shown in Table 1, Static LoRA and Residual MLP rely on a single adapter, limiting their ability to capture diverse historical residual patterns. Fixed-pool MoE introduces multiple experts but relies on a learned router for selection. In contrast, MemTime associates each residual expert with a memory prototype and retrieves relevant experts according to the current forecasting state, enabling more targeted corrections based on historical experiences.
  • Experimental Comparisons: We compare MemTime with Static LoRA, Residual MLP, Fixed-pool MoE, and Frozen Mem Bank under the same backbone and data protocol. As shown in Table 2, MemTime achieves lower average MSE and MAE than all four baselines across all six datasets. On Exchange, it reduces average MSE by 13.3% over Static LoRA and 10.1% over Fixed-pool MoE, supporting the benefits of retrieving residual experts based on the current forecasting state. Its 7.4% reduction in average MSE over Frozen Mem Bank supports the contribution of dynamic memory maintenance in preserving useful historical corrections.

Q3 & Q4. Target-domain Supervision and Data Partitioning

We address Q3 and Q4 jointly to clarify the data usage and isolate the contributions of memory adaptation.

  • Clarification of Data Usage: The description in Appendix A.3 was inaccurate and led to confusion regarding our data usage. In the actual implementation, all offline training of auxiliary modules, including residual generation, clustering, expert fitting, and retrieval-network training, uses the training split, not the validation split. A separate subset of the training split is used for hyperparameter selection and threshold calibration.
  • Validation and Testing Protocol: The validation split is used only to evaluate online adaptation, with memory updates following the prediction–feedback–update protocol in chronological order. This separation allows us to assess online adaptation on subsequent data with fixed hyperparameters, without tuning them on the same evaluation stream. Before testing, we reload the training-derived checkpoint and initial memory bank, ensuring that validation-adapted states are not transferred to testing. Test-time updates strictly follow the same causal feedback protocol.
  • Fair Experimental Comparisons: To isolate the benefits of memory adaptation, we compare Static LoRA, Residual MLP, Fixed-pool MoE, and Frozen Mem Bank under matched backbone and data-access settings. These comparisons distinguish the benefits of adaptive residual correction from those of dynamic memory maintenance.

We have made every effort to address the reviewers’ concerns. Due to time constraints, the revised manuscript is still being finalized and will be available soon.

[1] Im, Jeonghwan, and Hyuk-Yoon Kwon. Cosa: Context-aware output-space adapter for test-time adaptation in time series forecasting. ICLR 2026.

[2] Kim, HyunGi, et al. Battling the non-stationarity in time series forecasting via test-time adaptation. AAAI 2025.

[3] Lee, Thomas L., et al. Lightweight online adaption for time series foundation model forecasts. ICML 2025.

[4] Solano-Carrillo, Edgardo, Sreerag V Naveenachandran, and Julia Niebling. TEDM: Time Series Forecasting with Elucidated Diffusion Models. ICLR 2026.

[5] Zhou, Shiqiao, et al. Stationarity-Aware Retrieval-Augmented Time Series Forecasting. KDD 2026.

Reviewer_jieFOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MemTime, a hippocampus-inspired memory framework for adapting frozen time series foundation models to evolving data distributions. The framework consists of three components: a reinforcement learning-based Cognitive Sentinel for selective memory invocation, a Mental Retracer for memory-driven forecasting trajectory correction, and a Memory Metabolism mechanism for maintaining useful experiences over time. Experiments on seven real-world forecasting datasets demonstrate promising improvements in forecasting accuracy and adaptation efficiency.

Critical strengths
  1. The paper addresses the challenge of adapting pretrained time series foundation models to non-stationary environments without updating their backbone parameters. Leveraging evolving memory to incorporate deployment-time experiences is a meaningful and practically relevant direction.
  2. MemTime integrates adaptive memory invocation, memory-based residual correction, and long-term memory maintenance into a unified framework. The Cognitive Sentinel and Mental Retracer provide complementary mechanisms for determining when and how historical experiences should influence forecasting trajectories.
  3. In particular, MemTime achieves substantial improvements on the non-stationary Exchange dataset. The ablation studies, parameter sensitivity analyses, and memory visualization experiments further support the effectiveness of the proposed components.
Critical weaknesses
  1. Although the paper considers test-time adaptation methods in its efficiency analysis, the main forecasting results primarily compare MemTime against frozen foundation models. Since MemTime trains auxiliary modules using target-domain validation data, comparisons with adaptation baselines under equivalent data access would provide stronger evidence for the proposed memory mechanism.
  2. The methodology is developed around flow-matching dynamics, and the primary implementation uses Sundial as the frozen backbone. It remains unclear whether the proposed memory adaptation mechanism can be directly applied to foundation models with different forecasting architectures.
  3. Some descriptions of the method could more clearly distinguish between frozen backbone parameters and trainable auxiliary modules.
  4. A few minor typographical and formatting issues remain, such as the caption of Figure 5 and inconsistent capitalization of MemTime.
  5. Figure 1 could provide a more detailed illustration of how retrieved residual corrections interact with the flow-matching trajectory.
Questions for authors
  1. How are the contributions of individual memory traces evaluated when multiple memories are retrieved simultaneously? Does the Memory Metabolism mechanism distinguish the improvement contributed by each trace, or does it estimate utility from the overall forecasting improvement?
  2. What is the computational overhead of creating and updating memory traces during streaming inference? Does the reported latency include the asynchronous optimization of newly generated residual experts?
4 comments
AuthorsOpenReview ↗

Response to Reviewer jieF [PART 1/3]

Many thanks to Reviewer jieF for providing a detailed and in-depth review. Our responses to your concerns are listed as follows.

W1. Insufficient Comparisons with Adaptation Baselines under Equivalent Data Access

We address the concerns regarding data partitioning and experimental comparisons as follows. Tables 1 and 2 in Part 2 summarize the method designs and forecasting results, respectively.

  • Data Partitioning: We clarify that the training split is used to train the auxiliary modules and initialize the memory pool. A separate subset of the training split is used for hyperparameter selection and threshold calibration. The validation split is used only to evaluate online adaptation under fixed hyperparameters. Before testing, we restore the training-derived checkpoint and initial memory bank to prevent transfer of validation-adapted states. We will clarify this protocol in Appendix A.3.
  • Experimental Comparisons: We add comparisons with additional adaptation baselines under the same training data and online feedback settings. Table 1 summarizes the key design differences among these methods, while Table 2 presents their forecasting performance. MemTime consistently outperforms these baselines, highlighting the advantages of combining prototype-based retrieval with dynamic memory metabolism.
AuthorsOpenReview ↗

Response to Reviewer jieF [PART 2/3]

Table 1: Method Design

MethodMulti-Expert PoolRetrieval MechanismMemory Metabolism
Static LoRA✗ (Single Expert)✗✗
Residual MLP✗ (Single MLP)✗✗
Fixed-pool MoE✓ (Fixed Capacity)Learned Router✗
Frozen Mem Bank✓ (Fixed Capacity)Prototype Similarity✗
MemTime (Ours)✓ (Fixed Capacity)Prototype Similarity✓ (Utility + Lifecycle)

Table 2: Comparison with Static LoRA, Residual MLP, Fixed-pool MoE and Frozen Mem Bank

DatasetPredMemTime (MSE/MAE)Sundial (MSE/MAE)Static LoRA (MSE/MAE)Residual MLP (MSE/MAE)Fixed-pool MoE (MSE/MAE)Frozen Mem Bank (MSE/MAE)
ETTm1960.293 / 0.3400.292 / 0.3420.294 / 0.3410.295 / 0.3420.296 / 0.3430.296 / 0.343
1920.328 / 0.3710.337 / 0.3760.333 / 0.3740.334 / 0.3750.331 / 0.3730.331 / 0.373
3360.368 / 0.4040.370 / 0.4070.369 / 0.4070.371 / 0.4060.370 / 0.4050.370 / 0.405
7200.407 / 0.4250.418 / 0.4330.413 / 0.4290.415 / 0.4310.411 / 0.4280.413 / 0.428
Avg0.349 / 0.3850.354 / 0.3900.352 / 0.3880.354 / 0.3890.352 / 0.3870.353 / 0.387
ETTm2960.174 / 0.2600.178 / 0.2600.176 / 0.2620.175 / 0.2610.177 / 0.2650.176 / 0.267
1920.229 / 0.2910.235 / 0.3040.232 / 0.2970.233 / 0.2990.231 / 0.2950.230 / 0.295
3360.270 / 0.3320.287 / 0.3420.279 / 0.3370.281 / 0.3390.276 / 0.3350.275 / 0.335
7200.336 / 0.3700.360 / 0.3900.349 / 0.3810.353 / 0.3840.345 / 0.3780.343 / 0.377
Avg0.252 / 0.3130.265 / 0.3240.259 / 0.3190.261 / 0.3210.257 / 0.3180.256 / 0.319
ETTh1960.335 / 0.3750.341 / 0.3810.338 / 0.3780.337 / 0.3770.340 / 0.3790.340 / 0.379
1920.379 / 0.4010.381 / 0.4080.380 / 0.4040.379 / 0.4030.380 / 0.4040.381 / 0.406
3360.408 / 0.4170.405 / 0.4240.409 / 0.4200.410 / 0.4220.408 / 0.4210.409 / 0.421
7200.424 / 0.4500.433 / 0.4580.429 / 0.4540.431 / 0.4560.428 / 0.4530.428 / 0.453
Avg0.387 / 0.4110.390 / 0.4180.389 / 0.4140.389 / 0.4150.389 / 0.4140.390 / 0.415
ETTh2960.267 / 0.3230.272 / 0.3320.270 / 0.3290.269 / 0.3270.271 / 0.3300.270 / 0.328
1920.327 / 0.3660.329 / 0.3740.328 / 0.3700.327 / 0.3690.329 / 0.3710.329 / 0.370
3360.354 / 0.3920.357 / 0.3990.356 / 0.3960.355 / 0.3950.354 / 0.3940.356 / 0.396
7200.386 / 0.4480.401 / 0.4420.394 / 0.4490.397 / 0.4500.391 / 0.4480.390 / 0.449
Avg0.334 / 0.3820.340 / 0.3870.337 / 0.3860.337 / 0.3850.336 / 0.3860.336 / 0.386
Weather960.160 / 0.2230.158 / 0.2060.161 / 0.2240.162 / 0.2250.161 / 0.2240.160 / 0.224
1920.207 / 0.2510.205 / 0.2530.208 / 0.2520.209 / 0.2540.208 / 0.2520.207 / 0.253
3360.252 / 0.2980.254 / 0.2990.255 / 0.3000.256 / 0.3010.255 / 0.3000.254 / 0.300
7200.315 / 0.3360.322 / 0.3400.319 / 0.3380.321 / 0.3390.318 / 0.3370.318 / 0.339
Avg0.234 / 0.2770.235 / 0.2750.236 / 0.2790.237 / 0.2800.236 / 0.2780.235 / 0.279
Exchange960.147 / 0.2050.237 / 0.2690.185 / 0.2390.190 / 0.2430.174 / 0.2280.168 / 0.224
1920.230 / 0.3750.362 / 0.4150.290 / 0.3920.298 / 0.3970.274 / 0.3890.260 / 0.385
3360.422 / 0.5090.558 / 0.5980.478 / 0.5510.492 / 0.5610.460 / 0.5390.450 / 0.533
7200.562 / 0.6520.683 / 0.7820.616 / 0.7140.630 / 0.7280.602 / 0.7010.590 / 0.690
Avg0.340 / 0.4350.460 / 0.5160.392 / 0.4740.403 / 0.4820.378 / 0.4640.367 / 0.458
AuthorsOpenReview ↗

Response to Reviewer jieF [PART 3/3]

W2. Generalization to Different Forecasting Architectures

We clarify the applicability of MemTime to different forecasting architectures as follows.

  • Framework Applicability: MemTime is not limited to flow-matching models. The Cognitive Sentinel and Memory Metabolism can be applied to different forecasting architectures for memory invocation and maintenance.
  • Adaptation to Different Architectures: The Mental Retracer needs to be adjusted according to the backbone architecture. While the current implementation generates residual velocity corrections for flow-matching models, it can be adapted to autoregressive models by correcting hidden representations or forecasting outputs.

W3–W5. Methodological Clarification and Presentation

We have revised the methodology description and figures to improve clarity.

  • Model Parameters: We clarify that the pretrained backbone remains frozen, while the auxiliary modules are trained offline and memory states are updated during streaming inference.
  • Formatting: We have corrected the caption of Figure 5 and unified the capitalization of MemTime throughout the manuscript.
  • Figure 1: We have revised Figure 1 to illustrate how retrieved residual corrections are combined with the frozen velocity field and integrated into the flow-matching trajectory.

Q1. Utility Estimation of Multiple Retrieved Memory Traces

We clarify the utility estimation of multiple retrieved memory traces as follows.

  • Memory Aggregation: When multiple memory traces are retrieved, their residual corrections are combined according to their retrieval weights. During ODE integration, the contribution weight wkw_k of each trace is accumulated.
  • Utility Estimation: Once the ground-truth target becomes available, the overall forecasting improvement is calculated as ΔL=Luncorrected−Lcorrected\Delta L=L_{\mathrm{uncorrected}}-L_{\mathrm{corrected}}. Each trace's utility is then updated according to its accumulated contribution wkΔLw_k\Delta L. Thus, utility reflects its weighted contribution to the overall correction rather than an independently measured causal improvement.

Q2. Computational Overhead of Online Memory Maintenance

We further evaluate the computational overhead of online memory maintenance and its impact on inference latency.

  • Memory Maintenance Cost: Over 5,000 streaming requests, MemTime creates 100 new traces, each with 10 optimization steps. Total memory maintenance takes 8 seconds, corresponding to an amortized cost of 1.6 ms per request.
  • Inference Latency: The originally reported latency excludes asynchronous optimization of newly generated experts. To evaluate its actual impact, we measure inference latency with and without background maintenance. On a single RTX 4090, mean latency increases from 25.6 ms to 26.3 ms, with peak GPU memory usage of about 1.65 GB.

These results show that asynchronous memory maintenance introduces limited additional overhead without significantly affecting streaming inference.

Due to the limited time, the revised manuscript will be coming soon.

Thank you for your diligent response. I have updated my rating.

Reviewer_n4uzOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MemTime, a hippocampus-inspired framework for adapting frozen time-series foundation models to non-stationary streaming environments without updating backbone parameters. It consists of three components: a Cognitive Sentinel that determines when memory-based correction is needed, a Mental Retracer that retrieves historical residual experts to correct forecasting trajectories, and a Memory Metabolism mechanism that consolidates useful memories and removes outdated ones.

The authors evaluate MemTime on seven real-world time-series benchmarks and report improvements over existing foundation models, with up to 26.03% MSE reduction on the Exchange dataset.Additional ablation studies, long-horizon evaluations, and efficiency analyses are provided to demonstrate the effectiveness of the proposed framework.

Critical strengths

1. A coherent memory-based adaptation framework. The combination of selective memory invocation, residual trajectory correction, and memory consolidation provides a reasonable approach to continual adaptation while keeping the backbone frozen. Separating the decision of when to adapt from how to adapt is also a meaningful design choice.

2. Evaluation of long-term stability and efficiency. Beyond forecasting accuracy, the authors investigate long-horizon performance and computational overhead. The results suggest that MemTime can maintain relatively stable memory consumption while introducing less than 5% inference latency overhead compared with the frozen backbone.

Critical weaknesses

1. Insufficient comparisons with existing adaptation methods. The experimental evaluation mainly compares MemTime with zero-shot time-series foundation models, while comparisons with existing online adaptation and test-time adaptation methods are largely missing. Since MemTime is specifically designed to enable continual adaptation under non-stationary streaming environments, comparisons with these adaptive methods would be more relevant and necessary to demonstrate its effectiveness. The lack of such comparisons makes it difficult to assess the actual advantages of MemTime over existing adaptation approaches.

2. Incomplete evaluation of computational overhead. The efficiency analysis mainly focuses on inference latency and memory consumption, but does not sufficiently account for the computational costs of training the auxiliary modules and maintaining the memory pool during streaming. In particular, the creation and optimization of new memory traces may introduce additional overhead. A more comprehensive evaluation of the overall computational cost would be helpful to better understand the practical efficiency of MemTime.

Questions for authors
  1. Could the authors include existing online adaptation and test-time adaptation methods as additional baselines for comparison? If not possible, please explain clearly which such comparison is not possible or not applicable.

  2. Can the authors report the computational costs of offline auxiliary-module training and online memory maintenance, including the creation and optimization of new memory traces?  How significant are these costs compared with the reported inference overhead?

2 comments
AuthorsOpenReview ↗

Response to Reviewer n4uz [PART 1/2]

Many thanks for the detailed and constructive comments. We have been trying our best to address your concerns. Please see the response as follows.

W1&Q1. Comparisons with Existing Adaptation Methods

We conduct new experiments to compare MemTime with representative test-time adaptation methods, i.e., COSA[1] and TAFAS[2], and an online adaptation method, i.e., ELF[3]. To comprehensively show the effectiveness of MemTime, we further include two more recent baselines, a diffusion-based forecasting model TEDM [4], and a stationarity-aware retrieval-augmented forecasting model SARAF [5]. As shown in Table 1, MemTime consistently outperforms the additional baselines. Unlike COSA, TAFAS, and ELF, which rely on post-hoc adjustments, SARAF, which uses discrete retrieval, and TEDM, which employs static diffusion, MemTime directly corrects forecasting trajectories during generation. By identifying and correcting errors early, MemTime mitigates trajectory drift and error accumulation over long prediction horizons.

Table 1:Comparisons with Existing Adaptation Methods and Recent Baselines

DatasetPredMemTime (MSE/MAE)COSA (MSE/MAE)TAFAS (MSE/MAE)ELF (MSE/MAE)TEDM (MSE/MAE)SARAF (MSE/MAE)
ETTm1960.293 / 0.3400.296 / 0.3430.305 / 0.3550.299 / 0.3460.304 / 0.3510.297 / 0.349
1920.328 / 0.3710.334 / 0.3750.345 / 0.3880.338 / 0.3790.332 / 0.3740.331 / 0.378
3360.368 / 0.4040.374 / 0.4080.385 / 0.4200.378 / 0.4120.372 / 0.4080.383 / 0.414
7200.407 / 0.4250.420 / 0.4350.437 / 0.4500.425 / 0.4400.412 / 0.4260.412 / 0.428
Avg0.349 / 0.3850.356 / 0.3900.368 / 0.4030.360 / 0.3940.355 / 0.3900.356 / 0.392
ETTm2960.174 / 0.2600.175 / 0.2610.182 / 0.2700.178 / 0.2640.173 / 0.2590.178 / 0.263
1920.229 / 0.2910.232 / 0.2980.242 / 0.3120.236 / 0.3020.228 / 0.2970.232 / 0.292
3360.270 / 0.3320.282 / 0.3380.295 / 0.3520.288 / 0.3420.276 / 0.3350.277 / 0.366
7200.336 / 0.3700.359 / 0.3850.381 / 0.4050.370 / 0.3900.373 / 0.4010.347 / 0.385
Avg0.252 / 0.3130.262 / 0.3210.275 / 0.3350.268 / 0.3250.263 / 0.3230.259 / 0.327
ETTh1960.335 / 0.3750.338 / 0.3780.350 / 0.3900.342 / 0.3820.368 / 0.3980.375 / 0.402
1920.379 / 0.4010.383 / 0.4060.395 / 0.4180.387 / 0.4100.405 / 0.4210.407 / 0.422
3360.408 / 0.4170.412 / 0.4220.425 / 0.4350.416 / 0.4260.439 / 0.4450.438 / 0.451
7200.424 / 0.4500.435 / 0.4560.450 / 0.4700.439 / 0.4600.469 / 0.4790.454 / 0.476
Avg0.387 / 0.4110.392 / 0.4160.405 / 0.4280.396 / 0.4200.420 / 0.4360.419 / 0.438
ETTh2960.267 / 0.3230.270 / 0.3280.282 / 0.3400.275 / 0.3330.279 / 0.3480.272 / 0.336
1920.327 / 0.3660.331 / 0.3710.345 / 0.3850.335 / 0.3750.349 / 0.3960.331 / 0.380
3360.354 / 0.3920.362 / 0.3980.375 / 0.4120.365 / 0.4020.377 / 0.4270.360 / 0.410
7200.386 / 0.4480.395 / 0.4490.418 / 0.4650.405 / 0.4550.438 / 0.4750.401 / 0.469
Avg0.334 / 0.3820.340 / 0.3870.355 / 0.4010.345 / 0.3910.361 / 0.4120.341 / 0.399
ECL960.148 / 0.2300.139 / 0.2270.149 / 0.2350.142 / 0.2280.141 / 0.2310.157 / 0.235
1920.164 / 0.2630.162 / 0.2580.165 / 0.2640.162 / 0.2560.160 / 0.2580.172 / 0.267
3360.174 / 0.2710.185 / 0.2800.195 / 0.2920.188 / 0.2850.187 / 0.2830.198 / 0.293
7200.215 / 0.3070.245 / 0.3250.260 / 0.3490.250 / 0.3360.242 / 0.3280.255 / 0.341
Avg0.175 / 0.2680.183 / 0.2730.192 / 0.2850.186 / 0.2760.183 / 0.2750.196 / 0.284
Weather960.160 / 0.2230.156 / 0.2210.165 / 0.2250.157 / 0.2110.160 / 0.2220.159 / 0.219
1920.207 / 0.2510.198 / 0.2450.212 / 0.2640.213 / 0.2580.210 / 0.2540.208 / 0.256
3360.252 / 0.2980.258 / 0.3020.270 / 0.3150.262 / 0.3080.261 / 0.3050.263 / 0.308
7200.315 / 0.3360.330 / 0.3450.354 / 0.3650.338 / 0.3520.335 / 0.3500.341 / 0.354
Avg0.234 / 0.2770.236 / 0.2780.250 / 0.2920.243 / 0.2820.242 / 0.2830.243 / 0.284
AuthorsOpenReview ↗

Response to Reviewer n4uz [PART 2/2]

W2 & Q2. Computational Overhead of MemTime

We understand your concern regarding the computational efficiency of MemTime and address it from two perspectives.

  • Computational Overhead: Table 2 provides a comprehensive evaluation of the training and memory maintenance costs. The results show that memory maintenance introduces limited additional overhead, with only a minor impact on inference latency and GPU memory consumption. In particular, the low amortized maintenance cost demonstrates that continuous memory updates remain computationally efficient in streaming scenarios.
  • Difference from Conventional TTA: In essence, the paradigm of MemTime is different from TTA. To decouple the computational overhead and reduce the risk of forgetting, MemTime transforms "synchronous peripheral parameter update" into "asynchronous internal dynamic intervention". Synchronous backpropagation can block the inference pipeline, resulting in delay fluctuations, while repeated parameter updates may cause catastrophic forgetting. MemTime mitigates these risks through asynchronous background optimization and independent memory pools, while keeping the backbone frozen.

Table 2: Computational Cost and Efficiency Profiling of MemTime.

ModuleBaseline (Sundial)MemTime (Ours)Notes
Offline Auxiliary TrainingNot required15–35 min/dataset; 12–16 GB/GPU on 4 × RTX 4090Backbone remains frozen. Includes auxiliary-module training, memory-bank initialization, and Sentinel GRPO rollouts.
Online Memory Maintenance CostNot applicable8 s total; 1.6 ms/request (amortized)100 new traces over 5,000 requests, with 10 optimization steps per trace. Maintenance runs asynchronously.
Inference Latency24.5 ms/step25.6 ms/step+1.1 ms (+4.5%) compared with frozen Sundial.
VRAM Footprint1,450 MB1,485 MB+35 MB (+2.4%) compared with frozen Sundial; fixed-capacity memory pool.

Due to time constraints, the revised manuscript is not yet ready but will be available shortly.

[1] Im, Jeonghwan, and Hyuk-Yoon Kwon. Cosa: Context-aware output-space adapter for test-time adaptation in time series forecasting. ICLR 2026.

[2] Kim, HyunGi, et al. Battling the non-stationarity in time series forecasting via test-time adaptation. AAAI 2025.

[3] Lee, Thomas L., et al. Lightweight online adaption for time series foundation model forecasts. ICML 2025.

[4] Solano-Carrillo, Edgardo, Sreerag V Naveenachandran, and Julia Niebling. TEDM: Time Series Forecasting with Elucidated Diffusion Models. ICLR 2026.

[5] Zhou, Shiqiao, et al. Stationarity-Aware Retrieval-Augmented Time Series Forecasting. KDD 2026.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Kindly Request for Discussion

Dear Chairs and Reviewers,

Thank you very much for your thoughtful and constructive feedback, which has been invaluable in strengthening our work. As the discussion period ends soon, we would appreciate your thoughts on our responses and additional experiments. We hope these have addressed your concerns and would be grateful if you could reconsider your assessment and score in light of the new evidence. We remain eager to have any further discussions.

Sincerely,

Authors.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.