CloudWatch Anomaly Detection Train On Longer Period of Time

Question

I have a need to create an anomaly detection alarm on a metric with months of data and I want to make sure it will catch both short term spikes and long-term negative trends in the data based on all the data we collect. The data is not very granular and is only published on an hourly and daily basis. CloudWatch document states:

The algorithm trains on up to two weeks of metric data

and

CloudWatch anomaly detection continually evaluates the model and makes adjustments to it to ensure that it is as accurate as possible. This includes re-training the model to adjust if the metric values evolve over time or have sudden changes

Source

Does this mean that the model will only look at the past 2 weeks of data at any point to detect anomalies, or will the initially training look at the past 2 weeks and then continue learning as it gets more data points?