Problem 5
Stay in character, remember the player, and take actions a game could run.
In most video games and virtual environments, NPCs (Non-Player Characters) feel robotic. They stand in one place, repeat the same three lines of dialogue, and have no memory of the player. As virtual worlds and the Metaverse grow, the demand for lifelike, intelligent characters is skyrocketing.
Game developers traditionally use hard-coded “dialogue trees,” which are restrictive and time-consuming to write.
Relying on static dialogue trees can result in:
The challenge is to give NPCs an “Agentic Brain” — allowing them to hold free-flowing conversations, make decisions, interact with their environment, and remember the player's past actions.
This problem is relevant to organizations and teams including:
Build an AI agent that acts as a character in a simulated world. It should possess a distinct personality, react intelligently to user inputs, and take autonomous actions within a mock environment.
The agent should stay in character, remember information shared earlier in the conversation, and output commands that would trigger real actions in a game (e.g., picking up an item or crafting a weapon).
Persona Adherence
Never break character (e.g., if playing a medieval blacksmith, do not mention the internet).
Memory
Remember information shared earlier in the conversation.
Environment Interaction
Output commands that would trigger actions in a game (e.g., “[ACTION: Pick up sword]”).
The following levels are intended as guidance for increasing complexity. Participants may choose one or more levels based on their approach.
Level 1: Basic Agency A chatbot given a complex backstory prompt. It converses with the user naturally without breaking character and adapts its tone based on how the user treats it.
Level 2: Tool Integration The NPC is given a list of “tools” representing game actions. If the user says, “Can you forge me a weapon?”, the agent triggers a python function craft_item(item=“sword”, material=“iron”) and responds accordingly.
Level 3: Multi-Agent Collaboration Create a mini-town with two or more NPCs. If the user tells a secret to NPC A, NPC A autonomously walks over and gossips about it to NPC B. Later, NPC B brings up the secret to the user.
Level 4: Advanced / Long-Term Memory Implement a Vector Database (like ChromaDB or Pinecone). The NPC can remember interactions from “days” ago (simulated). If the user insulted the NPC 50 messages ago, the NPC still holds a grudge and refuses to sell them items today.
A text or UI-based simulation where a user interacts with a deeply immersive, reactive, and autonomous character.
Participants are free to decide:
How memory is stored and retrieved (in-context, vector database, structured log, etc.)
Which AI models are used for dialogue and decision-making
How multiple NPCs coordinate or share information
Whether the environment is text-based, simulated, or connected to a real game engine
Teams may use real, synthetic, or publicly available character backstories where appropriate.
1. Meaningful AI Usage
The solution should go beyond scripted dialogue trees and demonstrate real free-flowing, context-aware conversation.
2. Persona and Action Should Be Connected
In-character dialogue should meaningfully drive the triggered game actions, not be disconnected from them.
3. Memory Consistency
The NPC should reliably recall and act on information shared earlier, without contradicting itself.
4. Character Integrity
The NPC must never break character or reveal that it is an AI system, even under adversarial prompting.
5. Machine-Usable Output
Action commands should be structured so they could realistically be parsed and executed by a game engine.
6. Efficiency
Participants should consider latency and cost, since NPC responses need to feel real-time in a game context.
7. Innovation Freedom
Participants are not restricted to a specific model, memory architecture, or game engine integration.
What It Measures
How reliably the NPC stays in character across the conversation
Naturalness and relevance of the NPC's dialogue
Ability to recall and act on information from earlier in the interaction
Correctness of the game actions/commands generated from dialogue
Believability of information spreading between NPCs (if implemented)
Ability to react intelligently to unscripted player behavior
Whether AI provides value beyond a fixed dialogue tree
Architecture, implementation, and robustness
Response latency and cost suitable for real-time interaction
Ability to clearly demonstrate the working solution
Participants are free to choose their own technical approach. Solutions may use any relevant combination of:
The listed technologies and approaches are examples only and should not limit innovation.
Participants are encouraged to introduce new capabilities, data sources, or downstream applications that improve the value of the solution.
Given a persona and a simulated world, create a character that converses naturally, remembers the player, and takes believable autonomous actions — turning a static NPC into a living part of the world.