We describe a new framework for self-modifying programs, that is programs which can execute what they have themselves written. On the first hand, we use an abstract machine which makes explicit some typical behavior, such as turning data into executable code and vice versa. Moreover memory is also separated between data (what we can read and write) and code (what we can execute). On the other hand, we add another level of granularity in memory location, in order to deal with the problem of interpreting intensional behavior of the program more easily.