Pharmacodynamics of Context: How Wave-Resonant Retrieval Amplifies Bias in Skewed Memory Distributions

In a companion paper (Paper 12), we established that context quantity in retrieval-augmented generation follows a dose-response curve with model-specific therapeutic windows. This paper addresses the complementary question: does context selection method matter? We introduce wave-resonant retrieval, a four-dimensional scoring function combining intensity, temporal recurrence, phase coherence, and anti-saturation. On balanced data, resonance improves diversity by 24%. On production data (3,740 episodic memories, 78% concentrated in two emotion categories), resonance catastrophically fails: 0 wins and 140 losses across two independent implementations (GPT-5.4 and Claude Opus 4.6). We identify the mechanism—phase coherence amplifies the dominant class, creating a resonance monopoly—and extend the pharmacokinetic framework with a pharmacodynamic dimension: even at the optimal dose, the wrong selection method converts medicine into poison.