- 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
-
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.
-
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
Response to Reviewer SAMU [PART 2/3]
Table 1: Performance Comparison with TMoE, MemCast, and STanHop
| Dataset | Pred | MemTime (MSE/MAE) | Sundial (MSE/MAE) | TMoE (MSE/MAE) | MemCast (MSE/MAE) | STanHop (MSE/MAE) |
|---|---|---|---|---|---|---|
| ETTm1 | 96 | 0.293 / 0.340 | 0.292 / 0.342 | 0.313 / 0.353 | 0.354 / 0.375 | 0.296 / 0.362 |
| 192 | 0.328 / 0.371 | 0.337 / 0.376 | 0.362 / 0.382 | 0.392 / 0.393 | 0.355 / 0.395 | |
| 336 | 0.368 / 0.404 | 0.370 / 0.407 | 0.391 / 0.412 | 0.425 / 0.414 | 0.385 / 0.427 | |
| 720 | 0.407 / 0.425 | 0.418 / 0.433 | 0.452 / 0.443 | 0.486 / 0.448 | 0.453 / 0.460 | |
| Avg | 0.349 / 0.385 | 0.354 / 0.390 | 0.380 / 0.398 | 0.414 / 0.408 | 0.372 / 0.411 | |
| ETTm2 | 96 | 0.174 / 0.260 | 0.178 / 0.260 | 0.179 / 0.262 | 0.183 / 0.266 | 0.177 / 0.266 |
| 192 | 0.229 / 0.291 | 0.235 / 0.304 | 0.232 / 0.297 | 0.247 / 0.305 | 0.243 / 0.308 | |
| 336 | 0.270 / 0.332 | 0.287 / 0.342 | 0.293 / 0.334 | 0.307 / 0.342 | 0.302 / 0.346 | |
| 720 | 0.336 / 0.370 | 0.360 / 0.390 | 0.394 / 0.394 | 0.407 / 0.398 | 0.402 / 0.404 | |
| Avg | 0.252 / 0.313 | 0.265 / 0.324 | 0.275 / 0.322 | 0.286 / 0.328 | 0.281 / 0.331 | |
| ETTh1 | 96 | 0.335 / 0.375 | 0.341 / 0.381 | 0.373 / 0.394 | 0.385 / 0.393 | 0.340 / 0.387 |
| 192 | 0.379 / 0.401 | 0.381 / 0.408 | 0.422 / 0.422 | 0.435 / 0.422 | 0.398 / 0.437 | |
| 336 | 0.408 / 0.417 | 0.405 / 0.424 | 0.465 / 0.448 | 0.475 / 0.444 | 0.450 / 0.472 | |
| 720 | 0.424 / 0.450 | 0.433 / 0.458 | 0.477 / 0.465 | 0.463 / 0.459 | 0.512 / 0.511 | |
| Avg | 0.387 / 0.411 | 0.390 / 0.418 | 0.434 / 0.432 | 0.440 / 0.430 | 0.425 / 0.452 | |
| ETTh2 | 96 | 0.267 / 0.323 | 0.272 / 0.332 | 0.285 / 0.336 | 0.290 / 0.339 | 0.300 / 0.353 |
| 192 | 0.327 / 0.366 | 0.329 / 0.374 | 0.362 / 0.391 | 0.377 / 0.391 | 0.391 / 0.408 | |
| 336 | 0.354 / 0.392 | 0.357 / 0.399 | 0.415 / 0.431 | 0.416 / 0.425 | 0.435 / 0.448 | |
| 720 | 0.386 / 0.448 | 0.401 / 0.442 | 0.426 / 0.449 | 0.417 / 0.446 | 0.440 / 0.462 | |
| Avg | 0.334 / 0.382 | 0.340 / 0.387 | 0.372 / 0.402 | 0.375 / 0.400 | 0.392 / 0.418 | |
| ECL | 96 | 0.148 / 0.230 | 0.134 / 0.231 | 0.145 / 0.245 | 0.199 / 0.276 | 0.157 / 0.259 |
| 192 | 0.164 / 0.263 | 0.154 / 0.251 | 0.163 / 0.262 | 0.198 / 0.277 | 0.227 / 0.304 | |
| 336 | 0.174 / 0.271 | 0.182 / 0.276 | 0.174 / 0.272 | 0.213 / 0.293 | 0.317 / 0.369 | |
| 720 | 0.215 / 0.307 | 0.220 / 0.311 | 0.217 / 0.311 | 0.254 / 0.325 | 0.435 / 0.447 | |
| Avg | 0.175 / 0.268 | 0.173 / 0.267 | 0.175 / 0.273 | 0.216 / 0.293 | 0.284 / 0.345 | |
| Weather | 96 | 0.160 / 0.223 | 0.158 / 0.206 | 0.153 / 0.228 | 0.165 / 0.212 | 0.263 / 0.302 |
| 192 | 0.207 / 0.251 | 0.205 / 0.253 | 0.201 / 0.255 | 0.217 / 0.296 | 0.332 / 0.393 | |
| 336 | 0.252 / 0.298 | 0.254 / 0.299 | 0.261 / 0.358 | 0.276 / 0.336 | 0.369 / 0.475 | |
| 720 | 0.315 / 0.336 | 0.322 / 0.340 | 0.347 / 0.345 | 0.339 / 0.380 | 0.456 / 0.526 | |
| Avg | 0.234 / 0.277 | 0.235 / 0.275 | 0.241 / 0.297 | 0.249 / 0.306 | 0.355 / 0.424 | |
| Exchange | 96 | 0.147 / 0.205 | 0.237 / 0.269 | 0.148 / 0.216 | 0.183 / 0.226 | 0.192 / 0.211 |
| 192 | 0.230 / 0.375 | 0.362 / 0.415 | 0.246 / 0.395 | 0.277 / 0.296 | 0.295 / 0.312 | |
| 336 | 0.422 / 0.509 | 0.558 / 0.598 | 0.393 / 0.457 | 0.653 / 0.523 | 0.446 / 0.524 | |
| 720 | 0.562 / 0.652 | 0.683 / 0.782 | 0.861 / 0.690 | 1.014 / 0.753 | 0.862 / 0.765 | |
| Avg | 0.340 / 0.435 | 0.460 / 0.516 | 0.412 / 0.440 | 0.532 / 0.450 | 0.449 / 0.453 |
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.
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.
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.
The differences across MemTime, MemCast, TMoE, and STanHop lie in methodology and performance perspectives, which can be seen as follows.