I read the paper, but I’m not sure that I understand how the memory unit works. It seems like they reserve special tokens to read and write to this memory betwe

I read the paper, but I’m not sure that I understand how the memory unit works. It seems like they reserve special tokens to read and write to this memory between layers, as guided by the self-attention heads. In this way, the model can learn to squirrel away important details from the input prompt for later recall in subsequent layers. A global chalkboard mechanism. Am I on the right track?