Game AI Agent Engine

Characters That
Truly Come Alive

LuckNum gives game studios an AI agent engine built on long-term memory, persona modelling and autonomous decision-making. NPCs remember every player, read the context, and act where no script was written — turning open worlds from authored pipelines into stories that grow.

Explore Technology View Architecture
Agent decision latency
20ms
Concurrent agents
100K
Persona consistency
96.4%
Flat vector illustration: a physics chain reaction — a ball rolls down a ramp into a domino run, and the last ball flies toward a bullseye target
Memory Recall12 ms
Decision Planning20 ms
Concurrent Agents100K

Vision

Games Begin When NPCs Stop Being Scripts

For two decades, in-game characters have been driven by state machines and behaviour trees — precise, predictable, and therefore lifeless. We believe that once agents have memory, persona and goals of their own, players no longer walk through a script: they enter a world that remembers them, answers them, and may even resist them.

01

Memory

A cross-session long-term memory graph that holds every choice, relationship and emotional turn — so that meeting again always has a past.

02

Persona

Stable persona modelling keeps character, tone, values and motives consistent across an entire live-service cycle.

03

Autonomy

Goal-driven autonomy: agents live, ally, clash and grow even with no player around, so the world keeps running on its own.

Flat vector illustration: six side-by-side physics sandboxes — a ramp and ball, a Newton cradle, dominoes, a spring oscillator, gear trains and a lever, linked by dotted lines
Agent simulation cluster · tens of millions of interaction turns per day

Technology

Six Capabilities That Make Characters Alive

Six technical pillars working in concert, forming an agent system that keeps evolving with your live service.

Long-Term Memory

Episodic, semantic and relational memory kept apart, decaying and reinforcing by importance — so an NPC still recalls that trade and that betrayal six months later.

  • Memory Graph
  • Vector Recall
  • Cross-Session

Persona & Emotion

Character, tone, verbal tics and values encoded as a verifiable persona vector, paired with an emotion state machine — no identity drift across long sessions.

  • Persona Embedding
  • Affect Model
  • Consistency

Goal-Driven Autonomy

Utility scoring, behaviour trees and LLM planning scheduled together: routine behaviour runs on cheap rules, pivotal decisions go to model reasoning.

  • Utility AI
  • HTN Planner
  • LLM Planner

Multi-Agent Society Simulation

Up to a hundred thousand agents populate an offline copy of your world, letting economy, factions and rumour evolve on their own — then feeding the settled society back to live servers.

  • Agent Society
  • Scalable Sim
  • Emergent Narrative

Match Data Flywheel

Live telemetry, human labelling and self-play form a closed loop: the more players play, the better agents read that kind of player — and the cheaper live content becomes.

  • Telemetry
  • Self-Play
  • RLHF

Low-Latency Inference

A dual-path runtime for server and client: dialogue and action share one timeline and resolve within 20 ms, so the player never feels the seam.

  • Quantization
  • Edge Runtime
  • 20 ms

Architecture

A Four-Layer Game Agent Stack

From engine integration to application delivery, every layer exposes standard interfaces that can be adopted or replaced independently — without touching your existing combat and quest systems.

L4

Applications

Applications

  • MMO
  • Open World
  • Competitive
  • UGC Platforms
L3

Agents

Agents

  • Persona
  • Memory Graph
  • Goal Planning
  • Affect & Relations
L2

Models

Models

  • Dialogue Foundation
  • Behaviour Policy
  • Voice & Expression
  • World Knowledge
L1

Engine Integration

Engine Integration

  • Unity SDK
  • Unreal Plugin
  • Server Authority
  • Inference Runtime

Platform

Three Product Lines, One Agent Foundation

Flat vector illustration: a gear-driven core driving a driven wheel by belt, surrounded by icon-based capability modules
LUQ-Mind

Agent Core

A single runtime for memory, persona and goal planning, embedded into your existing server as an SDK — no rewrite of quests or scripting, and a first memory-bearing NPC running within half a day.

  • 3Memory Tiers
  • 20ms
  • 100K Agents
Flat vector illustration: a top-down sandbox with several physics rigs running in parallel, showing agents simulating different experiments at once
LUQ-Forge

Agent Studio

Configure personas visually, author goal functions, replay match traces and fine-tune online. Designers turn "this shopkeeper should be colder" into shipped behaviour without writing a line of code.

  • 0Lines of Code
  • 30Templates
  • 1Day Onboarding
Flat vector illustration: an open physics sandbox — a ball rolls down a ramp, crosses a truss bridge and slams into a crate stack, with a spring and wooded hills nearby
LUQ-Live

Agent Cloud

Elastic inference clusters and global memory storage support agents at a million-plus DAU, billed by real usage — so an empty server no longer costs you money during a content lull.

  • 100M DAU
  • 99.99%
  • 3Regions

Scenarios

From Scripted to Emergent, Across Six Genres

MMORPG

NPCs with memory and self-driven world events move long-term live content from hand-scheduled to emergent.

Open World

Townsfolk live out their routines, relationships and feuds on their own — the world keeps running after the player logs off.

Competitive Training

Human-like opponents and AI teammates scale to the player's skill, then hand back an actionable post-match review.

Emergent Narrative

Side quests and dialogue generated from player behaviour in real time replace fixed branching trees — the story no longer runs out.

Social Companion

A virtual companion with stable persona and emotion that remembers what you said last time — retention, not a gimmick.

Esports & Streaming

Automated casting, tactical breakdowns and human-versus-agent show matches — a directable roster of agents for events and streams.

Flat vector illustration: the break shot in pool — the cue strikes the white ball along a straight line, scattering the triangular rack as balls fly out
Multi-agent society simulation · 100,000 agents performing live

By the Numbers

Capability, Defined by Measurable Metrics

20ms Agent decision latency
100K Concurrent agents per cluster
96.4% Long-horizon persona consistency
8M Interaction turns generated daily
30Titles Games running on the engine
1.2B Dialogue foundation parameters

Milestones

Five Years, From Zero to One

  1. 01

    Founded

    LuckNum is founded; the team starts from multi-agent reinforcement learning in games and commits to a unified memory–persona–autonomy roadmap.

  2. 02

    First Agent Core

    The LUQ-Mind prototype is released, validating a thousand agents coexisting over long horizons — 30 days of uninterrupted operation with no memory corruption.

  3. 03

    First Commercial Launch

    The first partnered MMO ships with built-in AI NPCs, reaching 40× the daily player–NPC dialogue volume of the scripted era.

  4. 04

    Simulation Platform

    Daily generated interaction turns pass 8 million across combat, social and economic systems, with settled society fed back to live servers.

  5. 05

    Open Ecosystem

    LUQ-Forge enters public beta with 30 games on the engine, alongside a behaviour template marketplace and agent audit tooling.

  6. 06

    Client-Cloud Runtime

    A lightweight on-device runtime lands, resolving expressive dialogue and decisions within 20 ms on the client while the server keeps memory and authority checks.

Contact

Let's Make In-Game Characters Come Alive

Whether it is integrating the engine, technical collaboration or joining the team — we would love to hear from you.

Business [email protected] Research [email protected] Careers [email protected]