Water leakage through dam foundations is a critical problem for Algerian dams, particularly in arid regions where every cubic metre of storage is economically and socially significant. The Ouizert dam in Mascara province has been losing water through its left bank since impoundment in 1989, with an average annual loss of 10.42 hm³. The present study extends the monitoring record to April 2026, assembling a 37-year daily dataset of 12,521 validated measurements — the longest continuously analysed seepage record for any Algerian dam — together with piezometric data from 14 monitoring wells on the left bank. A two-step data-cleaning procedure combining physical bounds and Z-score filtering on annual regression residuals removed 0.6% of daily leakage records and fewer than 3% of piezometric readings, yielding a high-quality dataset for analysis and modelling. Four machine learning models were trained to predict monthly leakage from reservoir and piezometric inputs: linear regression, support vector regression, random forest, and gradient boosting. Lagged values of leakage itself were deliberately excluded from the feature set, requiring models to capture the physical link between reservoir state and seepage rather than exploiting autoregressive continuity. Gradient boosting performed best on the held-out test period 2019–2026 (R² = 0.668; RMSE = 86 L/s; MAE = 73 L/s), with random forest close behind (R² = 0.628). Both linear regression and support vector regression failed with strongly negative R² values, confirming that seepage at this site follows a non-linear law. Five-fold time-series cross-validation gave consistent results (gradient boosting: 0.678 ± 0.139; random forest: 0.674 ± 0.157). Feature importance analysis shows that three- and six-month moving averages of reservoir level dominate predictions, reflecting hydraulic memory timescales of weeks to months in the karstic foundation — a physically interpretable result consistent with piezometric response times documented at comparable dam sites. The maximum verified daily leakage in the cleaned dataset is 1,076 L/s. The 430 m NGA threshold identified in 2008 remains a robust operational marker after 25 years. The methodology and monitoring framework presented here are transferable to other Algerian dams with fractured-rock foundations.