Game AI
In one sentence: NPCs with a real cognitive architecture — they hold grudges, panic under stress, grow attached to the player, radicalize through 5 measurable stages, and can be de-radicalized. Not scripted, not prompted — emergent.
Theory sources: BM (G-profiles as personality, WM capture, trust hysteresis), NM (radicalization formula $R = Q/H \cdot (1-SIT)$, social bonds), EMT ($R_{expr}$, SIT, meme lifecycle), AGI_F (lifecycle phases, cognitive biases)
Implementation: Same bmc-core engine. Environment = game world. S-layer = game events.
The Problem
Current game AI is either scripted (FSM: if trust > 0.7: friendly) or prompted (LLM with a character description). Neither produces psychologically realistic characters:
| Situation | FSM/Script | LLM Prompt | BMC NPC |
|---|---|---|---|
| Player betrays NPC | trust -= 0.5 | “I am disappointed” (text) | Trust hysteresis: fast drop, slow recovery. GRIEF activated. “Traitor” meme with high fidelity. I-filter: future player words filtered through “traitor” hub |
| NPC under stress | if stress > 0.8: flee | “I am scared” (text) | FEAR captures WM ($k_{eff}=1$). Regression to simple strategies. May attack from fear (FEAR$\to$RAGE transition). Later remembers ($\kappa=2$) and changes attitude toward danger |
| NPC witnesses injustice | No reaction | “That’s unfair” (text) | SIT-gap: “why?” $\to$ SEEKING. If RAGE$\uparrow$: protests. If CARE$\uparrow$: helps victim. If FEAR$\uparrow$: stays silent but remembers ($\psi$) |
| Extended interaction | Repetitive dialogue | Varies text | Social bond grows. $R_{expr}$: NPC wants to share something important. Reveals secrets only when trust $> \theta$ |
What BMC Brings to NPCs
Personality from Architecture
Each NPC’s personality is a G-profile — 8 Panksepp drive weights:
- Hero: SEEKING$\uparrow$, FEAR$\downarrow$, CARE$\uparrow$, PLAY$\uparrow$ — risks life for others because CARE $>$ FEAR
- Villain: RAGE$\uparrow$, CARE$\downarrow$ — radicalizes through 5 measurable stages
- Merchant: SEEKING$\uparrow$ (resources), FEAR$\uparrow$ (cautious) — cheats when FEAR $<$ SEEKING (profit $>$ risk)
- Mentor: CARE$\uparrow\uparrow$, PLAY$\uparrow$ — teaches because CARE drives knowledge sharing
- Coward: FEAR$\uparrow\uparrow$, RAGE$\downarrow$ — but under extreme stress, FEAR$\to$RAGE transition (cornered animal)
These are not tags — they are computational parameters that determine behavior through the same cognitive cycle that produces emergent language and drug discovery.
Memory That Matters
$\kappa$-levels give NPCs real memory:
- $\kappa=0$: just saw it (sensory buffer, fades in seconds)
- $\kappa=1$: remembers (short-term, can be forgotten under stress)
- $\kappa=2$: will never forget (long-term, consolidated through sleep). The betrayal that shaped a character’s worldview
Villain Radicalization (5 Stages)
A BMC villain isn’t evil “because the script says so.” They radicalize through a measurable process:
$$R = \frac{Q}{H} \cdot (1 - SIT)$$- Trauma/Grievance $\to$ high SIT-gap $\to$ rumination
- Ideological offer $\to$ gap closure $\to$ VTA reward
- Hub displacement $\to$ ideology captures connections $\to$ diversity$\downarrow$
- I-recalibration $\to$ in-group/out-group hardening
- G-restructure $\to$ RAGE/FEAR$\uparrow$, CARE/PLAY$\downarrow$
The player can de-radicalize the villain: restore LP $> 0$, expose contradictions, reactivate PLAY. This isn’t a scripted redemption arc — it’s an emergent scenario with real cognitive mechanics.
Player Modeling
The same architecture models the player:
- SIT-profile: which puzzles “hook” the player (high SIT)
- G-profile: exploration (SEEKING) vs combat (RAGE/FEAR) vs social (CARE)
- Flow optimization: $SIT_{bore} < SIT < SIT_{anxiety}$ $\to$ adaptive difficulty
- $\kappa$-levels: what the player remembers of the lore ($\kappa=2$) vs what they’ve forgotten ($\kappa \to 0$) $\to$ NPC reminds contextually
Technical Advantages
| Property | Value |
|---|---|
| Compute | $O(N \cdot K)$ per tick — runs on CPU, no GPU needed |
| Inspectable | Full meme graph: you can see why the NPC made that decision |
| Persistent | Checkpoint save/load. NPC state survives game sessions |
| No training | G-profile + environment = behavior. No dataset, no fine-tuning |
| Lightweight | ~50KB per agent state. 1000 NPCs $<$ 50MB |
Market
| Segment | Size | Notes |
|---|---|---|
| Game AI middleware | $1.5B+ (2024) | Growing with narrative-driven games |
| AAA game budgets | $50–300M per title | BioWare, CD Projekt, Larian — deep pockets, hungry for innovation |
| Narrative games | Fastest growing genre | Player demand for deeper NPC interactions |
What We’re Looking For
- Game studio interested in a pilot: one quest, one town, 10–50 BMC NPCs with emergent behavior
- Middleware partner to package BMC NPC engine as a plugin (Unity, Unreal)
- Narrative designer to explore emergent storytelling with cognitive NPCs
For the cognitive model underlying NPC personality, see Personality. For trust hysteresis and social bond mechanics, see Network Memetics. For the radicalization model, see AI Safety.