AI agents waste time retrieving old memories — new system makes them forget selectively like humans do

Researchers built a memory system for AI agents that deliberately forgets old information instead of keeping everything accessible at all times, reducing the computational slowdown that happens as conversation histories grow. In practice, this means AI agents can now work faster and more accurately on long tasks by deciding when to look things up, rather than drowning in constantly-retrieved irrelevant context.

Current AI agents running on language models treat memory like a filing cabinet where every drawer is always open — the more history they accumulate, the slower they get and the more confused they become from conflicting information. This paper shows that selectively making old memories harder to access (but not deleting them) lets agents stay responsive and focused while still preserving long-term patterns. The practical upside: AI systems that can actually handle long, complex tasks without degrading, which matters for anyone building agents that need to maintain strategy over hours or days of interaction.