There's a fundamental limit on the amount of NPCs, because every NPC has to keep information about every other NPC in the context of the LLM, which scales quadratically in complexity with every word of information added.
LLMs have a large amount of built-in knowledge in the form of statistical relationships, but it's static and can't be modified. LLMs also have a working memory called the context that you can modify. This is sometimes also called the prompt, but in the case of this game it's constantly being changed dynamically. The model loses coherence and starts taking up immense amounts of resources due to quadratic scaling if the context becomes too large. Some information can be stored outside of the context and only retrieved when necessary in a process called retrieval-augmented generation, but if you want the LLM to use information in an implicit way, such as naturally referencing other characters in your game, it has to exist in the context or it doesn't exist to the LLM at all.