Finish projects, not lessons.

Here is Monday's bot in round numbers, so you can follow the sums:

• a system prompt of 400 tokens, sent on every call
• each customer message: about 60 tokens
• each bot reply: about 120 tokens

Because the whole history goes back each time, the input on turn n is the system prompt, plus every earlier exchange, plus the new message:

input(n) = 400 + (60 + 120) × (n − 1) + 60

Turn 1 sends 460 input tokens. Turn 20 sends 3,880. If the model's window is 8,000 tokens, then somewhere after turn 40 the request plus the reply no longer fits, and your code has to start dropping the oldest turns. The order number from turn 1 is the first thing to go. That is the first complaint, and it isn't a memory failure. It's a budget overflow.

Now add up every turn's input across a 42-turn conversation. The intuitive guess, "42 messages at 60 tokens each", is about 2,500 tokens. The real figure is around 174,000, roughly seventy times more, because each turn pays again for everything before it. Cost grows with the square of the conversation's length, not with the number of messages. Traffic went up a fifth, conversations got longer, and the bill tripled. That is the second complaint, and it has the same root cause as the first.

The third, the invented return policy, needs a different tool: giving the model the real policy text to work from and checking its answer against it. That has a name, RAG, and a whole module of its own later.