Enhancing Pretrained Model-based Continual Representation Learning via Guided Random Projection — ThinkLLM