arXiv:2611.21648v1 Announce Type: new Abstract: Hoel (2021) proposed that dreams distort replayed experience so that the brain does not overfit to recent events. We test this hypothesis on an LLM agent (Claude Opus 5.5) that stores what it learns in a text memory and rewrites that memory each night from the day's conversation transcript. The agent goes through ten simulated days of conversation built from LongMemEval. We corrupt 5, 10, 20 or 40 percent of the transcript's turns before each nightly rewrite in one of two ways. Detail corruption changes one or two details in a turn to plausible wrong ones, or replaces the turn with one by the same speaker from an earlier day. Random-text corruption replaces half of the selected turns with a passage of Moby-Dick and inserts Moby-Dick words into the other half. A size limit on the memory, rising from 19,500 to 117,000 tokens over ten nights, set from human free-recall data, forces it to forget. We learned that corruption did not help the agent apply what it learned in earlier conversations to new tasks in any of 30 runs (ten conditions, three different conversation histories). Replacing details in 40 percent of turns lowered recall from 76 to 60 percent. Random text had no consistent effect on recall, with a paired difference of 0 $\pm$ 2 points at 40 percent, because the model recognized it as corrupted and left it out of memory. Replaced details were stored as fact even when another turn held the correct value, and the agent answered 40 percent of those questions correctly, compared with 75 percent without corruption. These results suggest that a pretrained model can separate corrupted text from real events, so a fair test needs a learner that cannot make this separation as easily, such as a model fine-tuned on the corrupted days or an agent whose days take place in an invented world.
Can an LLM dream? Testing the overfitted brain hypothesis on an agent with text memory
Corrupting up to 40% of a conversation transcript before an LLM agent's nightly memory rewrite did not help the agent apply earlier lessons to new tasks in any of 30 runs, according to arXiv paper 2611.21648v1, which tested Hoel's 2021 "overfitted brain" hypothesis on Claude Opus 5.5 across ten simulated days built from LongMemEval. Detail corruption that replaced facts in 40% of turns lowered recall from 76% to 60%, and the agent answered only 40% of questions about replaced details correctly versus 75% without corruption, while random Moby-Dick text had no consistent effect (paired difference of 0 ± 2 points at 40%) because the model recognized it as corrupted and excluded it from memory. The authors conclude a pretrained model can separate corrupted text from real events, so a fair test of the dream hypothesis needs a learner that cannot make that separation, such as a model fine-tuned on the corrupted days or an agent operating in an invented world.
Test memory-rewrite pipelines with detail-level corruption, not random noise: Claude Opus 5.5 stored plausible wrong details as fact even when a correct turn existed, so check recall against ground-truth turns.
more in playbook: Evals and testing for LLM appsRun your AI side-project on zahid.host
EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.