It comes at a cost of a fragmented memory space, which is fine for some applications, like LLMs