Agenda
How miHoYo Builds Many-to-Many AI NPCs for Social Gameplay in 'Petit Planet'
Tim Ding (Lead AI Designer for Varsapura, miHoYo)
Ningyuan Jiang (Lead AI Designer for Petit Planet, miHoYo)
Pass Type:
Academic Pass, Festival Pass, Game Changer Pass, Indie Pass, Start-Up Pass -
Get your pass now!
Hot Topics:
AI
Track:
Design, Programming
Audience Level:
Intermediate
Format:
Lecture
Vault Recording: TBD
Audience Level: Intermediate
Most AI NPC systems are built for one player, one character, one private conversation. Petit Planet puts that interaction inside a persistent 3D world shared by many players at once. That shift changes the architecture at every layer.
Petit Planet is a life-simulation game from miHoYo. Its central AI character, Nalo, runs continuously across shared servers, remembers individual players across sessions, and responds to whoever is nearby in real time.
This session covers three concrete design and engineering problems that make many-to-many AI different from a companion app or single-player NPC: attention modeling across simultaneous players, memory context management across distinct per-player relationship histories, and post-training for social credibility rather than factual accuracy. On the memory side alone, human long-term memory is estimated in the low megabytes. Nalo's runs in the tens to hundreds of terabytes across her full player base, with no fixed context window and no upper bound on player count.
We validated this system across two large-scale playtests with over 50,000 players. This talk covers what we built, what we learned, and what we would do differently.
Petit Planet is a life-simulation game from miHoYo. Its central AI character, Nalo, runs continuously across shared servers, remembers individual players across sessions, and responds to whoever is nearby in real time.
This session covers three concrete design and engineering problems that make many-to-many AI different from a companion app or single-player NPC: attention modeling across simultaneous players, memory context management across distinct per-player relationship histories, and post-training for social credibility rather than factual accuracy. On the memory side alone, human long-term memory is estimated in the low megabytes. Nalo's runs in the tens to hundreds of terabytes across her full player base, with no fixed context window and no upper bound on player count.
We validated this system across two large-scale playtests with over 50,000 players. This talk covers what we built, what we learned, and what we would do differently.
Takeaway
1. A concrete framework for the three core technical differences between 1-to-1 AI NPC systems and many-to-many social AI: attention modeling, memory context management, and community-scale post-training. Each requires a distinct engineering approach, and this talk explains why.
2. A practical walkthrough of a multi-layered LLM architecture built for persistent, cross-player, cross-character retention at live-service scale. This includes active and passive forgetting mechanisms, separate retrieval paths for different memory types, and how pre-seeded memory establishes stable character identity before first contact. Nalo's total memory footprint runs in the tens to hundreds of terabytes, and we will explain how we keep that coherent in production.
3. Lessons from two playtests with 50,000+ players globally: What kind of outcome can we observe, and what design and engineering decisions we would make differently from the start
2. A practical walkthrough of a multi-layered LLM architecture built for persistent, cross-player, cross-character retention at live-service scale. This includes active and passive forgetting mechanisms, separate retrieval paths for different memory types, and how pre-seeded memory establishes stable character identity before first contact. Nalo's total memory footprint runs in the tens to hundreds of terabytes, and we will explain how we keep that coherent in production.
3. Lessons from two playtests with 50,000+ players globally: What kind of outcome can we observe, and what design and engineering decisions we would make differently from the start
Intended Audience
AI engineers, gameplay engineers, and game designers working on or planning to build AI NPC systems in multiplayer or live-service games.