Errors are Essential for General Intelligence
One of the problems with traditional approaches to Artificial Intelligence is the general philosophy that considers human cognition suboptimal. It is true that humans are error-prone. We make poor decisions without fully considering and weighing all the facts, we don't plan things out beyond a few steps, we fail to recall long-known facts, and sometimes we even forget what we are doing at a given moment. These errors, however, are what makes us optimal, rather than sub-optimal, for the complex world in which we live.
First, we operate in a rich and uncertain environment, and trial-and-error methodology is unavoidable for fact gathering. Current AI approaches already acknowledge the superiority of trial-and-error approaches for unknown environments (Reinforcement Learning).
Second, our lifespans aren't infinite, and most decisions have to be made relatively quickly. In fact, it would be suboptimal to gather all relevant knowledge for decision-making, if by the time that all knowledge is gathered the decision point will have come and passed (H.Simon called speedy decision-making w/o all relevant consideration satisficing).
Even as scientists acknowledge these two points as to necessity of some human "error", they argue that there is no reason to hinder machines with other errors. Why should a machine forget where it put its proverbial glasses? Why curse machines with the "it's on the tip of my tongue" phenomenon?
Forgetting, however, is not suboptimal. Decaying the irrelevant is a way of highlighting the relevant. Our ability to operate in real time with incomplete information in the face of a constantly changing environment relies on the ability to have relevant information at the ready.
"Who is to say what information is relevant?", you say. Well -- you. You and your environment. Memory items associated with great emotional changes in you are likely to be relevant to you. Memory items that are prompted recently or frequently by the events in the environment are likely to be relevant again in this environment soon. Thus, these memory items will be highlighted in your memory at the cost of other ones.
Humans and animals display much greater general intelligence than AI. So, why not copy the mechanisms of biological cognition into our artificial systems?
And don't forget to include forgetting.