The Dawn of Digital Life.
Move beyond static 3D models and scripted dialogue trees. The Aeion Character Engine V2 is a fully procedural, biomechanical, and cognitive simulation. It generates hyper-realistic digital humans from mathematical first principles and powers them with a Phase 5 Cognitive Architecture—featuring episodic memory, reinforcement learning, and genuine Theory of Mind.
A Full Stack from Genetics to Cognition
Pure Procedural Geometry
No artist-authored meshes. Every vertex, follicle, and pore is generated mathematically from a genetic seed. Characters are born from DNA, resulting in infinite unique variations and perfect anatomical landmarks.
Biomechanical Muscle Control
Discard canned animation clips. The Motor Cortex natively coordinates over 800 simulated muscles utilizing Recurrent Neural Networks (RNNs), solving kinematics and force calculation at 120Hz for incredibly fluid, dynamic movement.
7-Layer PBR Skin & Life Signs
Characters breathe. The idle microdynamics engine simulates 4 autonomic life signs: blinking, breathing, micro-saccades, and pupil hippus (dilation). The 7-layer skin shader dynamically reacts to the character's internal state—blushing when embarrassed or sweating when exerting force.
Cognitive Architecture (The Brain)
The Phase 5 engine simulates 5 major brain regions. The Prefrontal Cortex handles Goal-Oriented Action Planning (GOAP), while the Limbic System processes real-time emotional valence and arousal, continuously dictating the character's actions and physiological reactions.
Episodic & Procedural Memory
Characters remember. The engine utilizes a 4-tier memory architecture (Working, Short-Term, Long-Term, and Procedural). Characters index personal experiences with emotional weighting, allowing them to recall traumatic or joyful interactions and adjust future behavior autonomously.
Social Intelligence & Theory of Mind
NPCs maintain complex mental models of other agents. They track relationship intimacy, dominance hierarchies, and can predict what other characters know or believe, allowing for deep, unscripted social contagion and cooperation.
Neural Voice & Procedural Lip-Sync
Native integration with ElevenLabs and Qwen3 TTS generates emotional speech, while the Neural Bus parses the audio waveform to drive 30 distinct procedural viseme blendshapes for flawless, instant audio-driven lip-syncing without baked animations.
Smart Object NLP Interactions
Characters understand the physical world via a Natural Language Processing (NLP) grammar engine. An agent or player can issue complex sequence commands ("grab the heaviest sword, wait 3 seconds, then drop it"), and the NLP engine mathematically resolves the superlatives and translates them into physical IK actions.
The Universal Event Stream
Dual-Channel Observability
Whether a character's emotion is changed by a visual designer dragging a slider in Motion Studio, or by an autonomous AI agent triggering the `spatial_npc_set_emotion` MCP tool, the OS fires a unified `spatial.npc.*` event.
Deterministic Recording
Because the Cognitive Architecture is fully event-driven, every thought, gaze, and physiological morph is recorded to a central event bus. You can replay entire sessions with perfect frame-by-frame fidelity.
Character Engine — Working Capability Surface
Character lifecycle
Top-level entity management — every character spawns, ticks, and disposes cleanly, so scenes with hundreds of comings and goings never leak memory or leave orphaned state behind.
Gesture playback
A full gesture library with emotional gating and interruption handling — a wave started mid-sentence yields gracefully to a more urgent gesture instead of glitching through both.
Humanoid mesh planning
LOD-aware skinned-mesh planning that keeps hundreds of NPCs on-screen without tanking frame rate.
Character registry
A per-scene registry with a spatial index, so proximity queries ("who's near the player") stay fast even in crowded scenes.
Save schema
A versioned save format with backwards-compatible field migration, so character state survives engine upgrades without breaking old saves.
Asset pipeline contract
A typed contract carries each character from genome to mesh to texture, so every stage of generation stays consistent and predictable.
Autonomic response system
The four life signs that make a character read as alive — blink, breathe, micro-saccades, and pupil hippus — running continuously in the background.
Pipeline orchestration
The full generation sequence — genome, mesh, skin, rig, cognitive bootstrap — runs as one coordinated pipeline instead of disconnected steps.
Emotion driver
The PAD model (Pleasure / Arousal / Dominance) that drives every character's expression and posture from a single emotional state.
NPC orchestration
Per-scene NPC lifecycle, LLM dialogue routing, and tick coordination, so dozens of characters stay in sync without stepping on each other.
Smart NPC brain
The GOAP planner, behavior tree, and per-NPC memory retrieval that decide what each character does next.
Personality-driven NPC factory
Instantiate game-grade NPCs straight from personality archetypes — no per-character animation or dialogue authoring required.
Emotion-motion bridge
Connects the emotion driver straight to the motion system, so a shift in mood shows up in posture and gesture automatically.
The 5-Region Brain — Phase 5 Cognitive Architecture
NPCs aren't dialogue trees. They're simulated minds running 5 distinct brain regions, each with characteristic processing properties:
- 1. Limbic System (Emotion Driver). Continuous PAD (Pleasure-Arousal-Dominance) state, updated every tick from sensory input + memory recall. Outputs the emotional baseline that everything else reacts to. Direct line to the autonomic response system + facial morph engine — characters blush, sweat, pale, dilate pupils in response to felt state.
- 2. Sensory Cortex. Per-NPC awareness of the spatial world: vision cone (FOV + range + occlusion), hearing radius with sound-source attribution, proprioception of own body state, social proximity awareness. Sensory events feed directly into Limbic + Memory.
- 3. Prefrontal Cortex (GOAP planner). Goal-Oriented Action Planning. Each NPC has a stack of goals ranked by priority + emotional weight. Planner searches the action space for sequences satisfying the top goal, accounting for the NPC's current capabilities + world state. Replans on goal change or world disruption.
- 4. Memory System (4-tier). Working memory (active in current scene). Short-term (last 100 events, emotionally weighted). Long-term episodic (every meaningful event indexed with PAD + actors + place + time). Procedural memory (learned action sequences). The LLM dialogue layer queries this memory when generating speech — characters reference past interactions without scripting.
- 5. Social Cortex (Theory of Mind module). Each NPC maintains a mental model of every other NPC + player it's met: relationship intimacy, perceived dominance, what they know/believe, last interaction outcome. NPCs can predict + manipulate other NPCs' beliefs. Coalitions form. Reputations spread.
The 5 regions run on different cadences. Limbic + Sensory tick every frame (60 Hz). Prefrontal tick every 500 ms unless event-driven (sensory alert, goal change). Memory consolidation runs as a low-priority background job. Social Cortex updates on social interaction.
LOD Simulation — From Hero NPC to Crowd
| Tier | Tick rate | Mesh | Cognition | Memory | Use case |
|---|---|---|---|---|---|
| **Hero** | 60 Hz | Full PBR + skin shader | Full 5-region | Full 4-tier | Player conversation partner, story NPCs |
| **Crowd** | 30 Hz | LOD1 mesh + simplified skin | Limbic + GOAP | Short-term only | Background NPCs in player's awareness |
| **Ambient** | 10 Hz | LOD2 mesh + flat shader | Limbic + scripted | None | Distant NPCs in player's peripheral vision |
| **Sparse** | 1 Hz | Billboard sprite | None | None | Crowd density beyond player's interaction range |
| **Statistical** | varies | None | None | None | Population statistics simulated without instantiation |
Use Cases — Where Cognitive NPCs Actually Matter
Corporate Training Simulations
Healthcare practitioners practice difficult patient conversations with NPCs that have real fear, denial, anger responses. Conflict-de-escalation training for security officers. Sales-objection practice for new hires. Each NPC has a configurable personality archetype; the Cognitive Architecture handles the emotional logic.
Game NPCs
Open-world games with hundreds of NPCs that actually live in the world — schedule routines, social relationships, factional politics. Player actions ripple through the social cortex. Reputations spread.
LMS Tutors
Students interact with subject-matter NPCs (a "Roman general" for a history class, a "molecular biologist" for chemistry). The NPC has the personality + knowledge profile; the LMS module wires it into lesson plans + assessments.
Hospitality / Concierge Avatars
Virtual hotel concierges that remember guest preferences across stays. Procedural sympathy + apology when complaints arise. Authentic-feeling micro-interactions during check-in flow.
Sports / Esports Avatars
Coach NPCs that watch player performance, give context-appropriate feedback, develop their own "coaching style." Spectator avatars react to in-game events with believable emotional ranges.
Therapy + Mental Health Practice
Provider-overseen NPCs for exposure therapy + role-play scenarios. Always-available practice for socially-anxious patients. (HIPAA + therapeutic-context guardrails apply.)
Live Performance + Vtuber
A single performer drives a complex NPC chorus or band, with the Character Engine handling per-NPC autonomic variation + reactive crowd response.
Spatial Sales / Storefronts
WebXR retail avatars that demonstrate products + answer questions + maintain consistent brand-voice over thousands of interactions.
NPC Behaviors That Aren't Possible With Scripts
Subtitle: Specific examples of emergent behavior that the cognitive architecture enables — and the alternative would require thousands of lines of brittle conditional logic.
- Reputation propagation. Player insults NPC A. A tells their friend B about it (via the Social Cortex). When player later approaches B, B is colder — even if the player has never spoken to B before. The engine never had a hand-coded "if rude to A, B is colder" rule; this emerged from the Theory of Mind module + memory + emotional weighting.
- Mood-dependent memory. An NPC in a happy mood is more likely to recall positive interactions with the player. The same NPC in a fearful mood recalls past threats. Memory retrieval is mood-conditioned via the Limbic system biasing the indexed-by-emotion lookup.
- Body language that contradicts dialogue. A character says "I'm fine" while the autonomic response system renders elevated breathing rate + clenched-jaw morph + pupil constriction. Players read the body language correctly. The Limbic state says "stressed"; the dialogue model says "deflecting." Different outputs from the same state, by design.
- Goal-driven exploration. An NPC needs supplies. GOAP planner identifies: "go to market, buy supplies, return." If the market is closed (world state changed during planning), the planner re-runs — picks a different route, asks for help, finds a friend who has surplus, etc. The behavior tree doesn't enumerate every contingency; the planner discovers them.
- Faction politics. Two NPCs each track the player's reputation with their faction. NPC A grants access if rep is positive; B refuses if rep is negative. Player can be welcomed and shunned in the same village. Faction states emerge from individual NPC opinions, not a global flag.
- Group dynamics. A pack of 10 NPCs running boids-based flocking shows emergent alignment + cohesion. Layer on squad tactics for combat-coordinated behavior: flanking, suppressive fire, retreat coordination — emerging from individual decision-making, not central choreography.
Frequently Asked Questions
An LLM by itself is stateless + has no embodiment. The Character Engine wraps an LLM (or multiple) in: persistent emotional state, spatial awareness, biomechanical body, episodic memory of every prior interaction, social model of other agents, goal hierarchy. The LLM is one component; the engine is the integration that turns it into a believable presence in a 3D world.
Yes — Tier-3 (Ambient) and below run on mid-range mobile hardware. Hero NPCs require a discrete GPU for the 7-layer skin shader + 800-muscle motor cortex at 60 Hz. The LOD system handles graceful degradation: a player on a phone sees the same scene with Hero NPCs rendered at slightly lower fidelity, but all cognition runs server-side via the Realtime module.
Configurable per-tier. Hero NPCs typically use a frontier model (Claude / GPT-4) for dialogue; Crowd NPCs use a cheaper model (smaller GPT / Bedrock Titan / local Llama); Ambient NPCs use scripted templates. The Pro+ AI Router (`/architecture/ai-router-integration`) optimizes per-NPC cost vs quality, capping spend at a configurable per-scene budget.
Every cognitive decision writes to the deterministic event stream. Replay a session by replaying the events with the same random seed → identical character behavior. Useful for: debugging NPC misbehavior, comparing two LLM versions on the same scenario, certifying training-sim NPCs behave consistently for regulatory review.
Yes — opt-in. The Long-Term episodic store can be marked as "shared" across NPCs in the same faction. Player insults one merchant; all merchants in that faction know within minutes. Faction-shared memories use the same retrieval mechanism as individual memory, scoped by faction membership.
Multiple guardrails. (1) The LLM Dialogue Director enforces personality + scenario constraints in every prompt. (2) Pre-output filter rejects content that violates the scenario rating (G / PG / PG-13 / R configurable per scene). (3) Player-side opt-in for sensitive themes. (4) Enterprise tier supports BYO LLM with BAA + content policy customization.
Both. The cognitive architecture is genuine simulation — emergent emotional state + goal-driven action. The output is rendered via a high-fidelity animation pipeline (gestures, facial morphs, motor cortex). You can pause the simulation + scrub the animation timeline like a traditional rig, but the behavior itself isn't keyframed — it's computed.
~10 seconds for a basic NPC from an archetype (Scholar, Warrior, Diplomat, Merchant). ~5 minutes for a fully-customized character with bespoke personality + backstory + custom asset overrides. Production-grade story NPCs with hand-tuned dialogue style: a few hours of designer time. Mass crowd NPCs: instantiated by the thousands from procedural genome seeds in seconds.