Corrupted 'dreams' did not help a Claude Opus 5.5 memory agent

Hoel (2021) proposed that dreams distort replayed experience so that the brain does not overfit to recent events. A new arXiv preprint tests that idea on an LLM agent, Claude Opus 5.5, which 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. Before each nightly rewrite, the researchers corrupt 5, 10, 20 or 40 percent of the transcript's turns, 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. To force the agent to forget, its memory has a size limit that rises from 19,500 to 117,000 tokens over the ten nights, set from human free-recall data.

The main result is negative. 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).

The two kinds of corruption failed in different ways. 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 ± 2 points at 40 percent, because the model recognized it as corrupted and left it out of memory. Replaced details, by contrast, were stored as fact even when another turn held the correct value. On those questions the agent answered 40 percent correctly, compared with 75 percent without corruption.

The authors read this as a limit of the test, not a verdict on the hypothesis. Their 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. They name two candidates: a model fine-tuned on the corrupted days, or an agent whose days take place in an invented world.

Key facts

  • The test used Claude Opus 5.5 as an agent that rewrites a text memory each night from the day's conversation transcript, over ten simulated days built from LongMemEval.
  • Corruption did not help the agent apply earlier learning to new tasks in any of 30 runs (ten conditions, three conversation histories).
  • Replacing details in 40 percent of turns lowered recall from 76 to 60 percent; replaced details were stored as fact even when another turn held the correct value.
  • Moby-Dick random-text corruption had no consistent effect (0 ± 2 points at 40 percent) because the model recognized it as corrupted and left it out of memory.
  • The authors say a fair test needs a learner that cannot separate corrupted text from real events as easily, such as a fine-tuned model or an agent living in an invented world.

Why it matters

Hoel's overfitted brain hypothesis says dreams are useful because they distort replayed experience, which keeps the brain from overfitting to recent events. This paper asks whether a similar trick works for an LLM agent that keeps a text memory and consolidates it overnight. In this setup it did not: corrupting the nightly transcript never helped the agent apply earlier learning to new tasks.

Who it affects

The paper is most relevant to researchers studying memory and consolidation in LLM agents, and to people who draw on neuroscience ideas about sleep and dreaming. Builders of agents that rewrite their own text memory from conversation logs may also find the setup of interest, since it ties memory size limits to human free-recall data.

How to use it

There is nothing to install or run. The useful part is the experimental design: ten simulated days of conversation built from LongMemEval, a nightly memory rewrite, a memory size limit rising from 19,500 to 117,000 tokens over ten nights, and two corruption types applied to 5, 10, 20 or 40 percent of transcript turns. The authors also point to what a better test would use: a model fine-tuned on the corrupted days, or an agent whose days take place in an invented world.

How solid is it

This is an arXiv preprint. The headline finding is consistent across its 30 runs, but those runs come from three conversation histories and a single model, Claude Opus 5.5. The recall numbers given are specific: 76 to 60 percent for 40 percent detail corruption, and 0 ± 2 points for random text at 40 percent. The authors themselves word their conclusion as a suggestion, not a proof.

Risks and caveats

The result does not show that Hoel's hypothesis is wrong for brains. The authors say a pretrained model can apparently tell corrupted text from real events, which undercuts this agent as a test subject, and they propose fairer tests that were not run here. The source gives the recall drop only for 40 percent detail corruption, so it says nothing about how recall behaved at the lower corruption levels. Findings for one model should not be assumed to hold for others.

“These results suggest that a pretrained model can separate corrupted text from real events”

— Abstract, arXiv 2611.21648