Zero-shot prompts evaluate native memory by querying language models with isolated entity and topic framing while web search tools and external retrieval mechanisms remain strictly disabled. By analyzing responses across repeated independent runs, the system uncovers which brand names, product categories, and associations exist entirely within the model's static training weights. Additionally, specialized probes inspect token-level log probabilities and branching paths, quantifying the statistical confidence behind each recalled term. This establishes an ungrounded baseline that can later be contrasted against live search grounding to identify brand hallucinations, training data blind spots, and latent competitive positioning.