Multi-Turn Memory

Models don't remember anything

This is worth saying plainly: llama3.2 (or any model, local or cloud) has no memory of your last call by the time you make a new one. Every single call is stateless, it only ever sees what's in the messages list you send that time. "Conversation memory" isn't a feature the model has, it's a pattern you implement yourself: keep a growing list, and resend the whole thing on every turn.

Lesson 9's checkpoint built a fresh messages list on every question, which is why the assistant there couldn't have answered "what did I just ask you?", it genuinely had no way to know. This lesson fixes that by reusing one list across turns.

The code, piece by piece

messages: list[dict] = []

Declared once, outside the loop, and never replaced, only appended to.

def turn(user_text: str) -> None:
    messages.append({"role": "user", "content": user_text})
    response = ollama.chat(model=MODEL, messages=messages, options={"temperature": 0})
    messages.append({"role": "assistant", "content": response.message.content})

Each call to turn() does three things: adds the new user message to the running list, sends the entire list (not just the new message) to ollama.chat(), then appends the model's own reply back onto the same list. That last step is easy to forget and breaks everything if you do: without the assistant's own words in the history, the model has no record of what it already said, and can contradict or repeat itself on the next turn.

turn("My favorite color is teal. Remember that.")
turn("What is my favorite color?")

By the second call, messages holds four entries: the first question, the first answer, the second question, and (about to be added) the second answer. The model only "remembers" teal because that fact is sitting right there in the list it was just sent, not because of any memory on Ollama's side.

The real cost of this pattern

Every additional turn means a longer messages list, which means more tokens read on every subsequent call, even ones that don't need the early history. A ten-turn conversation resends all nine prior turns before generating the tenth answer. Lesson 19 covers what happens when that list grows past a model's context window; production systems often trim, summarize, or otherwise limit history rather than keeping it unbounded forever.