This Hacker News thread discusses the importance of context management for AI NPCs in games. The original article proposes feeding AI characters only relevant events, rather than the entire game universe. Commenters explore challenges like maintaining the illusion of ongoing NPC lives when the player is away, and the "jailbreaking" issue where AI NPCs can access knowledge outside their persona. Solutions include simulating the game world at varying fidelity levels (as in X4), generating events upon player re-entry, and cleverly using writing to account for gaps in NPC knowledge. Parallelization, action masking, and integrating embeddings for subjective elements are also suggested. Some users express concern about the computational cost of maintaining large contexts, and others mention that current LLMs may be too costly or unreliable. Overall, the thread highlights the complexities of creating believable and engaging AI NPCs that are grounded in their specific game world context.