This is a problem related to this specific model. NPCs have a long-term memory and a rolling short-term memory. The short-term memory is never cleared completely, since there was no need to do this with other models. Mistral Small specifically seems to degenerate over time because it has such great context understanding for its size, but isn't smart enough to self-correct like the much larger DeepSeek model. There is also something token sampling related in 1.5 that Mistral Small doesn't react well to.
Again, I recommend switching to Nemo or DeepSeek until 1.6.