Skip to main content
A world is the foundational unit of organization in the ecosystem. It provides an isolated container for an agent’s relationships, history, and facts.
The problem: Why AI forgetsTraditional AI models are largely hobbled; they operate effectively only within short, back-and-forth dialogues. Once that context window fills up or the chat ends, they lose the thread. Because they lack long-term memory, their interactions are fresh starts, making it impossible for agents to take on big, multi-day projects like a human coworker might.The solution: Linked memoryInstead of just hoping the AI finds the right data in a massive pile, Worlds creates a structured map, acting as a graph. By linking related facts together, your agent can always follow the logical path from one piece of information to another, no matter how much data you add.

​
Goals

GoalDescriptionConstraint
IsolationPrevent data leakage between agents.One RDF dataset per world.
PortabilityMaintain memory across model swaps.Decoupled from LLM provider.
FusionHybrid neural/symbolic reasoning.Requires both vector + SPARQL.

​
Architecture

A world functions as a neuro-symbolic memory container. It pairs an isolated RDF dataset with a high-performance vector search index.

​
Neural layer

The neural layer uses vector embeddings to manage semantic similarity. It allows the agent to navigate the world through natural language queries.

​
Symbolic layer

The symbolic layer uses RDF and SPARQL for deterministic facts. It provides deterministic validation, reducing the risk of hallucinations inherent in standalone LLMs.

​
Audit trails

Worlds eliminates opaque, untraceable behavior. Every retrieved fact has a verifiable audit trail in the RDF graph. If an agent makes a claim, the symbolic layer proves it.