NOAA’s National Weather Service (NWS) operates the Hydrologic Ensemble Forecasting System (HEFS), routinely providing probabilistic streamflow forecasts using meteorological forecasts from Global Ensemble Forecast System (GEFSv12). To develop a consistent set of ensemble streamflow hindcasts, the Office of Water Prediction has also run HEFS using GEFSv12 retrospective forecasts, creating an operational benchmark ensemble streamflow dataset that can be used to test changes and measure improvements. While HEFS uses a process-based Sacramento Soil Moisture Accounting (SacSMA) model, recent studies have shown that machine learning (ML) approaches are comparable and can often offer improvements over traditional hydrological models. The goal of this study is to use a data driven ML approach in the HEFS framework to forecast ensemble streamflow, thus creating a machine learning benchmark that can be directly compared to the existing operational benchmark. Specifically, we use a Long Short-term Memory (LSTM) network to train and test the forecast model at CONUS-wide priority watershed locations where OWP operational benchmark hindcasts are available. First, the LSTM model is trained using meteorological forcings from the Analysis of Record for Calibration (AORC), produced by the Office of Water Prediction (OWP)/ National Weather Service (NWS), with observed streamflows from United States Geological Survey as the target. Additionally, to capture basin-specific hydrological behavior, we incorporate basin characteristics from HydroATLAS - a global compendium of hydro-environmental attributes. The trained LSTM model is then tested using post-processed GEFSv12 re-forecasts - the same meteorological forcings used in the operational benchmark - enabling a consistent comparison between the LSTM model and the existing operational benchmarks. This framework offers insights into the potential of machine learning to complement and enhance the accuracy and skill of operational ensemble streamflow forecasting.