MemT-1.0: Memory-native foundation models MemTensor introduces a new architecture where persistent memory is part of model computation, enabling 1.86x inference speedup and 66x storage reduction at 128K history.

→ View original post on X — @_akhaliq