Chat memory and context limits

Learn how a chatbot 'remembers': the model itself has no memory, so your code sends past messages each time. Simple strategies: keep the last 4–6 turns; when the chat gets long, replace older turns with a short summary; always keep a token budget (system prompt + history + retrieved chunks + answer must fit the context window and your cost limit). Rewrite follow-up questions ('what about for part-time staff?') into full questions before searching.