This is the same as the paper about the little simulated town, where every character kept track of what happened each day. It provides context as to previous interactions, but without needing to keep the entire previous interaction in context.
Jonathan Stray's point of "I got the model to say a bad word" being the focal point of bias makes sense, and it's a lot easier to understand than the most insidious forms. Slurs are a starting point, but until we can easily point to "here are some situations where it went wrong," we're going to be falling back on the simple cases.
