WorldMind runs cognition for your NPCs: memory, opinions, and the hundred judgment calls a scene demands. Your engine still owns movement, physics, and rule enforcement. AI proposes. Your game decides.
Most game AI is brittle because behavior trees and state machines collapse once worlds become socially rich or tactically open-ended. WorldMind splits behavior into two layers. You keep authority over what matters. You offload only what hurts.
Millisecond loops, deterministic rules, authoritative physics and combat. Exactly how you wrote it.
The ambiguous calls. Who to trust, when to escalate, what lie to tell. The reactions that only make sense if the NPC remembers last week.
Each surface is opt-in and composable. Use one in isolation, or run a director over a fleet of embodied and tactical agents in the same world.
Quest givers, guards, merchants. They remember your name, your faction, and what you did last week.
Guard notices you now wear the rival faction's sigil. Warm greeting yesterday. Wary interrogation today.
Combatants, drivers, squadmates. AI plans 2–5 step sequences while your engine owns the twitch loop.
Police cruiser opts for a parallel route to A12, calls backup to B7, holds for trailing unit before engagement.
Colonists, villagers, sports leagues. Goals, moods, and grudges that persist from day 12 to day 400.
After the raid, Jana reprioritizes from hauling to medicine. She still resents Markus over last winter's theft.
Mission director, raid announcer, player copilot. Reads your UI, your minimap, your combat log, and the chat channel.
Copilot sees radar contact closing on your port quarter, suggests evasive burn and pre-charges shield capacitor.
WorldMind is a pipeline, not a black box. Sensors feed a context builder, memory services enrich retrieval, a planner proposes, a validator gates, and a router chooses the cheapest model that can hold the answer. Every stage emits traces.
After notable events, the runtime decides to ignore, remember, or promote. Ambient small talk decays. Betrayals persist. Repeated interactions strengthen confidence. Old episodes compress into summaries.
What happened recently. Time-stamped events.
Stable facts. Slowly earned, slowly forgotten.
Attitudes per actor. Deltas on every event.
Doctrine. Role instructions. Culture.
Working memory. Plans, threat list, conversation state.
Freeform text never drives your engine. WorldMind emits a validated structure with intent, ordered actions, and a mandatory fallback. Your game accepts fully, partially, rejects, or requests a repair.
{
"intent": "intercept",
"confidence": 0.82,
"goal": "stop_suspect_without_collision",
"next_actions": [
{
"action": "navigate_to_road_segment",
"args": { "segment_id": "A12" },
"why": "parallel route gives faster intercept path"
},
{
"action": "request_backup",
"args": { "location": "B7" },
"why": "roadblock succeeds with trailing cruiser"
}
],
"memory_writes": [
{ "type": "episodic", "text": "suspect fled toward bridge, third attempt this week" }
],
"fallback_if_unavailable": "continue_safe_pursuit"
}█WorldMind ships as a runtime core plus domain doctrine packs. Prebuilt for the verticals below, extensible for anything with a modding surface or engine integration.
Flight sims, stealth games, dating sims, MOBA coaches. If it has game state, WorldMind can reason over it.
Talk to us →Universal frameworks die before shipping. WorldMind starts with a concrete cognition loop inside one vertical, then generalizes into an SDK and a hosted service.
We're partnering with a small batch of studios and modders for the first wave. Tell us your world, your engine, and the one behavior you've never been able to ship.
WorldMind is an AI cognition runtime for games. It lets developers keep full control over movement and mechanics while giving NPCs memory, dialogue, tactics and context-aware decision-making through structured, inspectable AI.