What a conversation has to remember
A single message rarely carries all the information needed to respond well. A conversational system has to track what's already been said and established earlier in the same conversation, things like the user's name, a stated preference, or a partially completed task. This running record is called dialogue state, and it's what lets a bot say 'got it, I'll use your usual seat preference' instead of asking again for information already given three messages ago.
Losing track of dialogue state is one of the fastest ways to make a chatbot feel broken, even if each individual response is grammatically fine and on-topic. A bot that asks for a phone number the user already typed, or contradicts a price it quoted two turns earlier, isn't failing at language, it's failing at memory. Dialogue state doesn't have to be complicated: it can be as simple as a dictionary of key-value pairs updated after every turn, tracking things like the current intent, which slots are filled, and what the system last asked the user.