Tesseract improves after every recognition. How can I persist this improvements?

[Brais Gabín Moreira]

I'm running tesseract to read screenshots. I noticed that when I run the "easy screenshots" first and then the more difficult ones I get better recognition. This is not a feeling. I can reproduce this behaviour.

It's possible to export this new "knowledge" that tesseract learned to a new .traineddata?