This paper proposes the L2MP Framework — a Logic-Layered Modular and Memory-Preserving AI architecture — designed to address limitations in current LLMs such as lack of persistent memory, inability to self-learn, and weak prompt structure handling. It introduces multi-module routing for task segmentation (e.g., summarization, coding, reasoning), user-specific memory layers stored via vector databases, and ethical real-time web integration for continual learning. Experimental results demonstrate significant improvements in accuracy, relevance, and user satisfaction. L2MP aims to set a foundation for more intelligent, personalized, and autonomous AI assistants.