TEOM: Session-Based Virtual Embedding Matrices for Per-User Semantic Personalization in Large Language Models

We introduce TEOM (Token-level Ephemeral Overlay Matrix), a novel method for personalizing Large Language Model (LLM) behavior at the token level through session-based virtual embedding matrices. Unlike existing personalization approaches such as Soft Prompting, LoRA, or Knowledge Editing, TEOM creates an independent virtual embedding matrix that operates in parallel with the model's original embedding layer without modifying any base model parameters. Each user maintains a lightweight, portable profile that encodes their individual semantic preferences at the token level. The system supports three operating modes: full override, blended mode, and selective switching, controlled by a per-token transition coefficient σ. We demonstrate through proof-of-concept experiments that TEOM successfully shifts token semantics in sub-millisecond time while leaving all non-targeted tokens unaffected. The approach is ephemeral by default, interpretable, and requires no retraining, making it suitable for applications in personalized AI assistants, digital twins, cultural adaptation, and domain-specific customization.