Memory-Control Interaction and the Over-Intervention Effect in Memory-Augmented Language Models

This work investigates coherence degradation in memory-augmented large language models (LLMs) through a control-theoretic framework. We analyze the interaction between Cross-Session Narrative Memory (CSNM) and Entropy-Regulated Context Injection (ERCI) across controlled experimental conditions. We identify the "over-intervention effect", demonstrating that excessive feedback activation suppresses beneficial temporal structure introduced by memory, leading to reduced coherence and increased variance. We introduce the Self-State Vector (SSV), an adaptive control mechanism that dynamically regulates intervention based on system state. Experimental results show statistically significant improvements over static control (Cohen’s d = 0.588, p = 0.004), and strong robustness under adversarial perturbation. This work reframes LLM coherence as a dynamical systems problem governed by memory-control interaction, providing a foundation for principled design of memory-augmented language systems.