Persistent world infrastructure for games.
The world remembers because its state persists and moves. MistScale keeps your world’s identities, relationships, knowledge, and state alive between sessions, and AI Characters are the first product built on it.
How the platform is organized
Persistent world primitives describe your world. The platform foundation runs and connects them. Products are built on both.
Products
Built on the primitives and the foundation
Persistent world primitives
The foundation of your world’s state
Identity and relationships
Who is in your world, who knows whom, and how they feel about each other.
- Player and NPC identities
- Per-player trust, patience, and mood
- NPC-to-NPC trust
Persistent state
What each entity carries from one session to the next.
- Per-character memory
- Relationship state
- Identity state
Information flow
How what one entity knows reaches another.
- Sharing between characters
- Sharing presets and trust thresholds
Canon
What is true in your world.
- Per-NPC knowledge files
- Scene context
Platform foundation
Runs the primitives and connects your game
Control plane
Projects, team access, and keys.
- Projects
- Team invites
- API keys
Developer platform
How your game connects.
- Unity, Unreal, Godot, and Web SDKs
- REST and WebSocket APIs
- Dashboard and Playground
Products on the platform
Built on the same identities, relationships, and state.
MistScale AI
AI Characters
Characters that read and write persistent entity state: they remember each player, change with each relationship, and pass on what they learn.
- Per-player memory
- Mood and emotion
- Trait drift
- Grounding
- Scene context
- Voice
Model research
We are training our own model.
MistScale AI runs on a cognition layer we built: retrieval, grounding, relationship state, mood. A general model sits at the center of it today. Varen is ours: a language model fine-tuned for game cognition, so the character work lives in the model rather than in the prompt around it.
Working name
Varen
A language model fine-tuned for game cognition and in-world characters.
- Type
- Fine-tuned LLM
- Built for
- Game cognition and NPCs
In development
Built for the loop, not for chat
The job is not open conversation. It is reading persistent state, grounding an answer in it, holding character, and deciding what was worth remembering. That is what it is tuned on.
Grounding as an objective
Today a separate pipeline checks specific claims against memory, lore, and the character brief. We are training that behaviour in rather than bolting it on.
State as a native input
Per-player memory, trust, and mood are already structured on our side. A model trained on that shape does not need it flattened into a prompt first.
We are not publishing weights, evaluations, or a date. When there is something worth measuring, we will show the working.
One fact, many readers.
The state is the product. One moment in your game becomes a fact that more than one system reads. This is one example of state that persists and moves, not the whole platform.
One fact
A player burns the village mill.
It reaches MistScale through conversation. Sharing follows per-character rules and trust thresholds.
- 01
Entity state
The player boasts about it to the tavern keeper, who stores it as a memory about that player.
- 02
Relationships
The tavern keeper's trust and mood toward the player shift over time.
- 03
Information flow
The tavern keeper is set to Open and trusts the miller at 6 out of 10, so the miller receives the fact as unverified hearsay.
- 04
AI Characters
When the player next talks to the miller, he can bring up what he heard.