adjective · AI concept
Stateless
also called no memory
Every request starts from a blank slate. The AI does not remember you; the app just hands it the whole story again each time.
Having no stored state between requests, so each call must contain everything the model needs and conversation "memory" is something the calling application builds.
The model keeps nothing between calls. Anything that looks like memory is the application storing the history and sending it back in as input on the next request, which is why the context window limits it.
Real-life example
You tell a chat assistant your name, ask three more questions, and it still uses your name. It feels like it remembers. What actually happened is that the app sent your name along with every new question. Use the model directly with just the last question and it has no idea who you are.
Where the analogy breaks
-
It is not forgetting, because there was nothing stored to forget. What looks like remembering is the app resending old messages.
-
Memory features are the app's, not the model's. When a product says it remembers you, it saved notes somewhere and puts them in the input.
-
Resending costs something. The longer the conversation, the more every new request carries, which is why trimming, summarizing and caching exist.