When I do coding with AI, and expect to be busy for a while, I ask the AI to address me as Mr. Tinklewags with his responses. It’s like a canary in the mine. When he starts to forget to call me that, I know all the other important things in my original prompt are now forgotten, and I have to repeat it again.
AI forgetfulness is an influential piece of the puzzle when performing tasks with transformer-based AI. Vendors tout their ability for needle-in-a-haystack capability, but it keeps forgetting essential stuff, especially when running out of its context window. Then auto-compacting, and summarization happen.
New efforts to solve this forgetfulness is called continual learning, with Google on the forefront with an architecture that remembers information on multiple levels.
Or maybe LLMs are just too good at being a human. I forget people’s names all the time.
