← News

Agents & MCPSources: arXiv · 18 Nov 2026

Nightly memory corruption did not help an LLM agent, study finds

A preprint tests the overfitted brain hypothesis on a Claude Opus 5.5 agent with text memory, corrupting transcripts before nightly rewrites across ten simulated days.

In plain words

Researchers tested whether deliberately scrambling an AI agent's memories each night, like dreaming may do for the human brain, improves learning. They corrupted parts of a language model agent's daily conversation records before it rewrote its memory. Corruption never helped, and the model often filtered out or mishandled the noise.

What was announced

Building on Hoel's 2021 proposal that dreams distort replayed experience to prevent overfitting, the authors ran an LLM agent that stores knowledge in a text memory and rewrites it nightly from each day's conversation transcript. The agent lived ten simulated days built from LongMemEval, with a memory cap rising from 19,500 to 117,000 tokens, set from human free-recall data. Before each rewrite, the team corrupted 5, 10, 20 or 40 percent of transcript turns in two ways: swapping plausible wrong details or replacing turns with Moby-Dick passages and vocabulary.

Across 30 runs spanning ten conditions and three conversation histories, corruption never improved how the agent applied earlier learning to new tasks. Detail corruption at 40 percent of turns cut recall from 76 to 60 percent, and replaced details were stored as fact even when correct values appeared elsewhere, dropping related question accuracy from 75 to 40 percent. Random text had no consistent recall effect, a paired difference of 0 plus or minus 2 points at 40 percent, because the model recognized it as corrupted and excluded it. The authors conclude that a pretrained model can separate corrupted text from real events, so a fair test needs a learner that cannot, such as one fine-tuned on corrupted days or an agent in an invented world.

What it does not show

  • Evidence is limited to the paper's abstract; full methods and results are not available in the supplied passages.
  • Findings come from one agent model and simulated ten-day conversations, so results may not generalize to other systems or real-world use.
  • The authors themselves note the test may be unfair to the hypothesis because the pretrained model could identify corrupted text, and no independent replication is reported.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.