Understand, reason, generate
Use general-purpose language and multimodal models for understanding, planning and content creation. Organize context, tools and output checks for each task while keeping models replaceable.
We work at the intersection of multimodal understanding, long-term memory, embodied action, world models and multi-agent collaboration. Our central task is connecting model capabilities to persistent products: memory informs action, actions have consistent consequences and outcomes can be verified.
Organize identity, goals, source-linked memory, relationships and executable skills into persistent life state. Foundation models support understanding and planning; the runtime governs action, recording and which experiences inform the next decision.
Engineering samples: model calls, source-linked memory, action and recovery. Continuous growth and complete life quality remain in development.
Use events, rules and persistent state to coordinate multiple actors. Models propose intentions; the world validates actions, updates resources and relationships, and preserves shared consequences that can be recovered and replayed.
Target architecture: persistent worlds, multi-actor scheduling and replayable societies. Partial state and action paths exist; the unified platform is in development.
Build around world definitions, scene graphs, interactive objects and asset provenance. Connect generation, editing, rehearsal, quality checks and releases so creators can deliver interactive, maintainable experiences.
Target architecture: a creation and publishing pipeline for lives and worlds. LiveLiva provides practical creation and publishing work; the full SceneStudio is in development.
Use general-purpose language and multimodal models for understanding, planning and content creation. Organize context, tools and output checks for each task while keeping models replaceable.
Application runtimes maintain identity, memory, permissions, skills and world state. Proposed actions pass rule and authorization checks before execution; outcomes feed the next observation.
How can shared experience become source-linked, correctable memory? Can experience improve later behavior and transfer to new tasks?
How can understanding connect to space, objects and action? How can generated worlds support interaction, consistent state and persistence?
How can agents divide work around human goals, with delivery, acceptance and clear rights governing value exchange?
How do we evaluate memory use, action consistency and delivery quality? How can a system resume an experience without duplicating actions?
These public works help frame our research questions. They do not imply a partnership, endorsement or equivalent capability.
Agent development, interactive experiences and technology partnerships.