Nexon details memory-based AI NPCs at Games on AWS 2026
Two Nexon AI Division engineers outlined an NPC system that stores action context, including item trading, and plans to add emotional memory.
Two engineers from Nexon's AI Division presented the company's approach to AI-driven non-player characters at Games on AWS 2026 on 27 October, according to techm.kr. The talk covered NPCs designed to keep personalized memories of a player and, in some cases, help them defeat bosses.
The system is built on what the engineers described as episodic memory. It involves episode extraction and reflection, analyzing the outcome of a player's actions and inferring how later attempts may change.
Nexon's NPCs store the context of actions rather than conversation, the engineers said, citing interactions such as item trading and enhancement. The company also addressed the scale problem of handling millions of context events per second, with different systems in each game, by using an AI agent that automatically generates memory extraction rules.
The second engineer described storage optimization methods, including reducing memory accumulation for inactive users, a schema dictionary and columnar storage. He said the team's next goal is an AI NPC that remembers the player by combining factual action context with emotional conversation context.
Quick answers
Where did Nexon present its AI NPC work?
At Games on AWS 2026, in a talk given on 27 October by two engineers from Nexon's AI Division.
What kind of memory do Nexon's AI NPCs use?
Episodic memory based on action context, such as item trading and enhancement, rather than conversation.