The Personalization Paradox: When Adaptive Context Hurts Small Language Models

Personalization is widely assumed to improve language model performance -- more user context should yield better responses. We show this assumption is wrong for small models. Through a systematic ablation study of Fabrik-Codek, a personal cognitive architecture built around a 7B-parameter local model with six adaptive algorithms (Thompson Sampling, adaptive forgetting curves, graph-guided retrieval, incremental profiling, confidence-based retrieval stopping, and learned task routing), we evaluate 10 configurations across 65 benchmark cases on two tracks: generic coding tasks and domain-specific tasks targeting user expertise. The results reveal a personalization paradox: the full pipeline degrades generic task performance by 19.8% compared to the raw baseline (0.634 vs. 0.791), while the same pipeline minus graph retrieval (A3) nearly matches the raw baseline on domain tasks (0.861 vs. 0.865) and outperforms it on medium and hard cases. We diagnose three compounding bottlenecks: graph expansion queries that inject irrelevant neighborhood context, competence-based model escalation that routes to a larger but less focused model, and prompt bloat from accumulated personalization layers. A targeted mitigation recovers +12.8% on generic tasks while reducing latency by 67%, confirming the diagnosis. These findings establish that personalization for small models should be conditional: applied selectively based on topic match and competence level, not uniformly. The architecture, algorithms, and evaluation methodology are open source and validated with 1,000 unit tests.