Does AI also need sleep? Performance declines as conversations get longer

Carnegie Mellon researchers proposed a sleep-like memory method that pauses input and consolidates context with weights to help LLMs handle long conversations.

In experiments, the approach improved accuracy, especially on more complex tasks where performance usually drops over longer dialogues.

The work aims to reduce memory and context-window strain in long-horizon reasoning and make AI behavior more stable over extended sessions.