Stateless describes the fundamental property of language models that they retain no memory between separate calls. Each request is processed in isolation, with no awareness of previous interactions, unless the relevant context is explicitly supplied again as part of the input. The model does not remember you, your last question, or anything it told you before.
This statelessness is why external memory systems exist. A model on its own cannot maintain a conversation across turns, recall a user’s history, or build up knowledge over time — every call starts from a blank slate. Any sense of continuity must be created by the surrounding application, which feeds the necessary past context back into each request.
Vector databases are the standard solution. By storing embeddings of past conversations, facts, and documents, and retrieving the relevant ones to include in each new prompt, they give a stateless model the appearance of long-term memory. Understanding that the model is inherently stateless clarifies why retrieval, persistent memory, and context management are not optional extras but essential parts of building coherent AI applications.