A runtime learning layer for small LLMs that enables behavioral adaptation without model fine-tuning. The system learns behavioral rules (not facts) from feedback, stores them with confidence scores ...