There are a few things that might be happening here. It's possible that textcat.teach queues up questions about the same text, but with different labels. This is especially likely if your model is produces uncertain scores for those examples. It's also possible that your text example has a very small change, which causes our hashing function to consider them different.
Do you perhaps have some examples that you can share?
Thankyou Koaning! I will create some examples of this. I am reasonably sure the examples were not slightly different, but it could definitely be because the model was doing poorly on those examples.