Context rot — the degradation of LLM reasoning accuracy during extended conversations — is caused by contradiction accumulation, not context length. Even Google's 1M-token context window drops 47.8 percentage points under contradictions, while removing contradictions restores performance regardless of length. We propose a metabolic architecture inspired by human sleep that processes accumulated knowledge during idle time: detecting contradictions, resolving conflicts, and preserving both old and new claims as temporal pairs. Across 8 open-source models (8B–27B) and 11 paired comparisons over 180-turn dialogues, metabolism-enabled systems significantly outperform controls (sign test p = 0.0107). In a controlled three-condition experiment (n = 3 per condition), metabolism achieves 73.3% accuracy versus 21.1% without metabolism (Kruskal-Wallis p = 0.027). An unexpected observation reveals that metabolized systems exceed even the contradiction-free baseline (73.3% vs. 56.7%, p < 0.05), suggesting that contradiction pairs act as knowledge anchors preventing fact retrieval degradation.
Akihito Sunagawa (Mon,) studied this question.