This study presents a novel linear recursive model incorporating temperature effects on both conversion and dissipation to reconstruct climate-driven terrestrial water storage anomalies (TWSA). Using GRACE/GRACE-FO and meteorological data, the authors generate daily, high-resolution global reconstructions from 2002 to 2021. The proposed method demonstrates significantly faster parameter convergence and superior performance compared to existing datasets in a majority of river basins and land grids. By refining the analysis of precipitation's contribution to TWS, the approach offers improved insights for separating climatic and anthropogenic impacts on water resources. The paper demonstrates substantial work, supported by extensive experimental testing. The title could be improved to more accurately capture the paper's core contribution.

• The authors could provide the rationale for not using the CSR Mascon data product in this study.
• The GLDAS-2.2 data is also forced with the meteorological analysis fields from the operational European Centre for Medium-Range Weather Forecasts (ECMWF) Integrated Forecasting System. Due to the data agreement with ECMWF, this GLDAS-2.2 daily product does not include the meteorological forcing fields. The GLDAS-2.2 data are archived and distributed in NetCDF format. Thus, the authors should explain the differences between the two temperature datasets.

Minor comments:

• L50, Usage of ‘Eq’ is not defined before.
• L135-136, the citation format ‘(Lan and Wenke, 2022)’ is wrong. The same for ‘(Muñoz-Sabater et al., 2021; Muñoz Sabater, 2019)’.
• L179 and L185, abbreviations with ‘where’ at the beginning of a sentence are unnecessary.
• L680, ‘In some other basins’ to ‘In other basins’.

The study presents a four-parameter, linear recursive model that reconstructs daily terrestrial water storage anomalies (TWSA) at 0.5° resolution for 2002–2021 using precipitation and temperature. The authors claim higher computational efficiency and better basin-scale performance than that of previous statistical reconstructions. While the work is interesting and potentially useful for data-scarce regions, my major concerns focus on (i) the lack of daily-scale validation, (ii) the arbitrary 116-basin subdivision and its global mapping strategy, and (iii) the insufficient demonstration of added value against physically based assimilation products such as GLDAS CLSM. These issues must be addressed before the paper can be considered for publication. I recommend Major Revision.

Major comments

1. The model is trained on de-seasonalised and de-trended monthly TWSA from GRACE/GRACE-FO, and daily fields are produced but only monthly datasets are assessed. Although GRACE provides only monthly observations, the daily reconstruction should also be evaluated by comparison with other independent datasets such as model simulations and aridity indexes.
2. In the work, TWS is reconstructed on land grids, but why do you evaluate β at the basin scale? In my view, you should not use this method to reconstruct grid-scale TWS, because the increase in TWS may be greater than precipitation due to runoff processes, which can lead to β > 1. Actually, Zhong et al. (2025) also apply a similar method to reconstruct basin-scale TWS, which can ignores the impact of runoff.
3. I think you should divide the datasets into training periods and validation periods, as in machine learning, to avoid the over-fitting risk.
4. Parameters a, b, c, d are claimed to be "physically meaningful" but lack independent corroboration such as ET/P partitioning or field infiltration measurements. Therefore, you should add scatter plots of a vs. observed ET/P and d vs. estimated groundwater turnover time, and discuss sign mismatches.
5. As you mentioned, GLDAS CLSM (0.25°, assimilates GRACE) offers a quasi-independent reference. GLDAS provides reasonable TWS estimates. You should compare the results between the reconstruction and GLDAS, and illustrate the advantages of your method.

Minor comments

1. Line 233: add the missing equation number.
2. Line 333: "2.81 % of grids with NSE > 0.8" disagrees with the histogram.
3. In Figures 6 and 16 (b–j), the letters (a, b, c, etc.) in panel (a) are preferable to ID numbers.