Skip to main content
Worlds provides a structured framework for world memory. Instead of treating an agent’s context as a flat list of chat logs or disjointed text chunks, Worlds organizes information as a dynamic, queryable model of reality.

​
The Worlds pipeline

To understand how Worlds powers intelligent agents, you must understand the lifecycle of data moving through the platform.

​
Ingestion

Raw information enters the system from user chats, GitHub repositories, or PDFs. At this stage, the data remains unstructured human language.

​
Neuro-symbolic engine

The Worlds engine uses LLMs to extract meaning and items. It translates ambiguous language into structured triples (subject → predicate → object). These facts then merge into a world—an isolated container where the graph evolves through:
  • Updating conflicting facts.
  • Extending existing items with new context.
  • Inferring hidden relationships via symbolic reasoning.

​
Retrieval

When an agent needs context, it performs a hybrid search. This process mixes semantic vector similarity with deterministic graph traversal to pull a high-precision slice of reality directly into the context window.

​
Storage engine

To achieve both semantic flexibility and structural precision, the storage engine employs a hybrid strategy. Worlds utilizes an in-memory, WASM-compiled RDF store that supports SPARQL. The infrastructure supports any RDF store—including Apache Jena Fuseki or a local file system—that implements rdf-patch forward synchronization. n3 is the preferred store because it runs entirely within the JavaScript runtime, providing isolated, high-performance in-memory state.
  • Pre-loading: WASM modules are pre-loaded to ensure “warm” isolates.
  • Hydration: The SQLite “system of record” hydrates the graph state upon initialization.
  • Edge cache: Hot state persists in the edge cache between requests for millisecond read latency.

​
SQLite storage

The system utilizes a hybrid schema for persistence to avoid the overhead of general-purpose SPARQL engines on disk while maintaining semantic integrity.
  • triples table: Stores atomic units of knowledge (Subject, Predicate, Object).
  • chunks table: Stores overlapping text segments with vector embeddings targeting string literals, and ranks derived from triple data.
  • entity_types table: An optimized table for mapping entities to their rdf:type IRIs, enabling rapid structural filtering.
  • blobs table: Handles large-scale RDF data and file-based state.

​
Hybrid search and RRF

The platform utilizes Reciprocal Rank Fusion (RRF) to combine results from distinct indices into a single, unified relevance ranking:
  • Semantic search: Captures conceptual meaning using a vector index and high-dimensional embeddings, specifically 1536-dim.
  • Keyword search, via FTS5: Provides exact term matching using the BM25 ranking algorithm.
  • Graph context: Restricts search results based on structural RDF relationships using subject or predicate filters.
The fusion algorithm follows the industry-standard RRF formula: score=∑d∈D160+rank(d)score = \sum_{d \in D} \frac{1}{60 + rank(d)}

​
Technical specifications

For a deeper dive into the mathematical and philosophical foundations of the Worlds storage engine, refer to the Whitepaper.