Techniques are disclosed relating to controlling cache size and priority of data stored in the cache using machine learning techniques. A software cache may store data for a plurality of different user accounts using one or more hardware storage elements. In some embodiments, a machine learning module generates, based on access patterns to the software cache, a control value that specifies a size of the cache and generates time-to-live values for entries in the cache. In some embodiments, the system evicts data based on the time-to-live values. The disclosed techniques may reduce cache access times and/or improve cache hit rate.