dreamengine

WildReality DreamEngine is an open-source AI simulation engine designed for LLM-powered roleplay and world simulation. Built as an Electron NodeJS application, it goes beyond simple LLM chat clients by implementing a full game-state management system with persistent world logic. Key features include: - **Item Persistence**: Objects placed in the world remain in their locations across thousands of messages unless explicitly moved. - **Character Evolution**: Characters have complex bond systems that cause them to evolve based on interactions with the user and other characters, with fully customizable relationship dynamics. - **User as Character**: The user is treated as a character within the simulation, not given special privileges or omniscience. - **Story Rules**: World rules (e.g., no magic in sci-fi settings, physical limitations) are enforced, and characters abide by them. - **Dynamic Character Traits**: Characters are defined by traits, hidden mental/emotional states, and environmental triggers, all customizable via JavaScript scripting. - **Individual Perspectives**: Each character experiences the world differently, with their own knowledge, memory, and perception. - **No Omniscience**: Characters cannot know events they were not present for, including the user's actions. - **Dynamic Context Resizing**: The context window automatically adjusts as the story progresses, keeping narratives fresh and limiting repetition. - **Power Scaling**: A built-in and extendable power scaling system prevents unrealistic character dominance. - **Model Compatibility**: Works with any compatible underlying model (Llama, Mistral, GLM, etc.). - **NSFW Capable**: Supports explicit content with conditional relationship gating based on bond strength. - **Enforced World Layout**: Granular world definition including item positions, character postures, and physical properties. - **Volume and Weight**: Physical simulation where items have weight/volume and characters have strength/carrying capacity. The engine uses JavaScript scripting (sandboxed) to define characters, worlds, and items rather than traditional character cards. This scripting system is not attention-based and does not consume tokens, allowing for arbitrarily complex character behaviors. **Technical Requirements**: The engine uses additional LLM inference to track character states and world logic. For best results, it recommends models with at least 120B+ parameters running on hardware like an RTX 6000 Blackwell with 4-quant versions. Smaller models may result in state corruption or 'ghostly actions.' **Deployment Options**: - Electron App: `npm run start` - Web Version: `npm run web` (creates a single-user secure server accessible via browser) - Local Server: Compatible with DreamServer (https://github.com/onzag/de-server) for running the engine on custom hardware The project is currently marked as Work In Progress (WIP).