You build a tool that stuffs a customer's whole email history into the prompt before asking the model to draft a reply. It works on short histories but on long ones the model "forgets" the earliest emails. What's happening?