Cognitive Twin: Personality Transfer via Small-Model LoRA with Runtime Knowledge Graph Augmentation

We present the Cognitive Twin architecture, a three-component system that produces a faithful digital replica of a human operator's conversational persona without baking volatile domain knowledge into model weights. The architecture separates personality (a LoRA adapter trained on the operator's historical responses), knowledge (a live knowledge graph queried at inference time), and trajectory awareness (geometric scalars characterizing conversation dynamics). We find that a Qwen2.5-3B model with LoRA adapters on all 36 transformer layers produces more authentic personality transfer than a 7B model with adapters on 8 layers. The smaller model's weaker RLHF conditioning is easier to override, and full-layer coverage is more important than parameter count. Training on 2,923 examples extracted from 4,698 session files, we observe a 2.5:1 ratio of correction signals to affirmations in operator responses, confirming that persona data is dominated by directive rather than confirmatory interaction. DoRA (weight-decomposed LoRA) OOMs on 16GB Apple Silicon for 7B models, making standard LoRA with comprehensive layer coverage the practical optimum. At inference time, the cc-graph-kernel provides provenance-tracked knowledge slices from 71,130 live triples, and the Anticipation Geometry framework supplies 7 trajectory scalars that condition response style on conversation momentum. The full system runs on two Mac mini nodes (M2 + M4, 16GB each) connected via Thunderbolt, with the adapter served through MLX at sub-200ms latency. **Keywords:** cognitive twin, LoRA, personality transfer, knowledge graph, runtime augmentation, Apple Silicon, anticipation geometry, conversational AI