Good points about lack of “memory”. In practice, this can be easily solved by using moving averages. MAs of multiple periods carry “memory” of previous prices. Then, each MA value at time “t” can be used in traditional ML “stationary” models. So simple feature engineering solves the issues you are identifying.