Thanks for the feedback, and that's a great idea I should have done that! They are extracted and added each turn, all handled by the same middleware proxy that also handles the retrievals. I put the full code for that (memory_proxy.py) in the article.