For decades NPCs have been driven by rule‑based systems, with personalities and interactions hard‑coded before release. The emergence of large language models (LLMs) enables NPCs to converse freely, adapt to new situations, and exhibit behaviors not explicitly authored, sparking extensive academic and commercial interest. Yet the field lacks a unified overview to assess what has been built, how well it works, and at what cost. To address this, we performed a systematic mapping of 377 scholarly papers, selecting 42 for detailed analysis around one main research question and ten sub‑questions across four thematic axes. Our study provides the first quantitative portrait of component adoption, latency, safety coverage, and reproducibility: short‑term memory and world‑state grounding appear in roughly 60% and 69% of works, while long‑term memory (31%), reflection (38%) and retrieval‑augmented generation (14%) remain rare, and response times range from 2.6 to 15 seconds. We also identify six structural barriers to commercial adoption and offer evidence‑based recommendations for developers and researchers.\n\nReview