WP-001. Can an LLM dream?

Erik Hoel proposed that dreams distort replayed experience so the brain does not overfit to recent events. We tested this on an LLM agent that rewrites its text memory every night from the day's conversations, with part of each day corrupted before the rewrite. Changing details in 40 percent of chat messages lowered recall on LongMemEval questions from 76 to 60 percent. The model recognized inserted Moby-Dick text as corrupted and left it out of memory. Neither kind of corruption helped the agent apply what it learned to new tasks, so our results do not support Hoel's idea for an LLM agent.