The core problem

The model is stateless; the agent isn't.

An LLM remembers nothing between calls. Any memory an agent has is something your code stores and feeds back into the next prompt.