The paper is about a scalable alternative to $k$-order Markov process called k-LAMP. k-LAMP only needs $nnz(T)+k$ (where $T$ is the transition matrix and $nnz(T)$ is the number of nonzero entries in $T$) parameters, while $k$-order Markov process needs as many parameters as the number of paths of length $k+1$ in $T$. A generalized version called $k$-GLAMP is also suggested, it needs $k*nnz(T)+k$ parameters. An experimental comparison to Markov process and LSTM (Long-Short-Term-Memory) seems convincing. ### Typos: - page 5, Theorem 11: "\n this version.)"