Does an AI learn while you chat with it?

You tell an AI assistant: "I always drink my coffee with oat milk and no sugar." Three messages later you ask for a breakfast recipe, and the AI immediately takes your preference into account. It feels as if it just learned about you.

Yet when you open a fresh chat without saved memory, or when someone else asks the same model, it knows nothing about it. Did the model actually learn during your conversation, or did it just read along?

The difference comes down to two separate places:

1. Model weights (the locked vault): Billions of numerical parameters are tuned during training before release. Once the model is deployed for users, those weights are frozen. Typing your oat-milk message does not change a single weight inside the base model.
2. The context window (the open notepad on the desk): Every time you send a new message, the app resends the conversation history in the background. The model simply re-reads your earlier sentence from the notepad on the desk.

• OpenAI Platform Docs: Text Generation & Context Window
• Anthropic Docs: Context Windows

Why does a chatbot still remember certain facts across separate chats? Because the application around the model keeps its own notebook (Memory) or looks up passages in your files before answering (Retrieval-Augmented Generation, or RAG):

• Memory feature: The app stores a short note such as "User prefers oat milk without sugar" in a database and quietly places that slip of paper into the context window of future chats. Delete the entry in settings, and the knowledge disappears immediately.

• Document search (RAG): A search step retrieves relevant passages from your documents and places them onto the desk for that specific question. Here too, the base model's weights do not change.

• OpenAI Help Center: Memory FAQ

There is a second, separate timeline to keep in mind: while you are typing, the running model does not change. Depending on your plan and privacy settings, however, a provider may use stored conversations later to train future model versions in a subsequent training run.

On many consumer accounts, model training can be turned off in settings; on API connections and business tiers, training on customer data is excluded by default. Before pasting sensitive material into a chat, check three layers separately: what sits in the current context window, what is stored by the memory feature, and what your training settings allow for future model generations.

• OpenAI: Data Controls FAQ
• Anthropic Privacy Center: Training Policy