We use erase operation to remove information that is not required in the memory. After performing an erase operation, we will have a new updated memory matrix where some elements in the memory would be erased. How can we erase the value of a particular cell in the memory matrix? Here, we introduce one more vector called the erase vector, , of the same length as the weight vector, . The erase vector consist of values of 0s and 1s.

Okay. We have an erase vector. But how can we erase values and get the updated memory matrix? We multiply by our memory matrix at a previous step, , and get the updated memory matrix .

That is, .

But how does this work? Only if both weight and erase elements at an index i are 1, then that particular element in the memory will be set to 0—in other words, erased; otherwise, it will retain its own value. For example, look at the following diagram. First, we multiply the weight vector,, and erase vector, :

Then, we subtract 1 from it—that is, , and we get a new vector as shown in the following:

Next, we multiply with the memory matrix at the previous time step, , and get the updated memory matrix,: