A self-improving agent can keep its weights frozen and still get worse through the memories it learns to trust.

These agents store past episodes, score them with an LLM, and reuse them later as precedents without changing model weights.

Wrong episodes can still receive high self-scores; across the paper's factual banks, the tested models endorsed 31% to 54% of their own wrong answers as correct.

Once that score enters persistent memory, the mistake can influence future decisions instead of disappearing after one bad answer.

The authors call this the Echo Gap.

Stronger or different LLMs did not reliably repair it, because their grading errors often remained correlated with the original self-grading bias.

• arxiv. org/abs/2608.00017

Title: "Memory Reward Inflation in Self-Improving LLM Agents"