You have a great conversation with AI, open a new chat, and suddenly it remembers nothing you were talking about. That's because of something called statelessness.
At its core, AI just predicts words, and by default it doesn't even remember the last message you sent. The only reason it follows a conversation at all is that the entire chat history gets fed back into it on every single message. You don't see this in ChatGPT, Gemini, Grok, or Claude because the interface hides it for you. But it becomes obvious the moment you use AI through the API, where you have to resend the full conversation history yourself.
This isn't a bug, it's how AI is designed to work. And what about the "memory" toggle in ChatGPT and Claude? Behind the scenes, it's just feeding your other chats into the current one. The model doesn't remember on its own.
Why this matters: if you ever want to go past using AI as a better Google and start building real systems with it, you need to understand this core behavior so you can design around it.