A good example of this was the first Gemini model that allowed 1 million tokens, but would lose track of the conversation after a couple paragraphs.