No memory between conversations

When you close a conversation and start a new one, the model knows nothing about the previous exchange. Every new conversation begins from a blank context — the only thing present is whatever the application decides to include in the system prompt.

This surprises people. A model can feel remarkably personal and responsive within a conversation, building on earlier points, remembering what you mentioned. Then you start a new session and it has no idea who you are.

Some applications add memory on top of the model — saving summaries of past conversations and injecting them into new context windows. This creates a simulation of continuity. But it's the application doing that work, not the model itself. The model's own architecture has no persistent memory.

This is worth understanding not to feel tricked, but because it shapes what these tools are good at. They're excellent for sessions — a single piece of work, a specific problem, an extended conversation about one thing. They're not inherently personal in the way a long-term assistant relationship might be, unless memory is added externally.