Skip to main content

​
At a glance

​
Overlap and shared capabilities

Both Worlds and Graphiti store facts as a graph, preserve provenance to source material, and retrieve through more than one signal. Both reject the idea that vector similarity alone is enough for agent context: structure decides what an agent can verify, not just what looks similar. The difference is in how facts enter the graph and what guarantees they carry:
  • Graphiti extracts entities and relationships from raw episodes with LLM calls at write time. Each edge carries a bi-temporal validity window, so the graph answers both “what is true” and “what was true on a date”. Conflicting facts are invalidated, not deleted.
  • Worlds treats ingestion as curation. Facts enter a world deliberately, through import and patches, as W3C RDF triples in an append-only ledger. Every quad records where it came from and when it entered.
Graphiti’s temporal windows solve fact churn inside one deployment. Worlds’ append-only ledger plus SPARQL gives you standards-based verification: the same graph queries against any backend, auditable provenance per quad, and no extraction step that can silently misread a source.

​
When Worlds fits

  • You need deterministic graph querying (SPARQL) to verify multi-hop relationships before an agent acts.
  • You require quad-level provenance and audit trails for every assertion.
  • You need portable context across backends and providers, without an extraction pipeline tied to specific models.

​
When Graphiti fits

  • Your facts change constantly and you need point-in-time answers (“what was true last quarter”) from automatically ingested conversations and events.
  • You want memory writes driven by LLM extraction rather than manual curation.
  • You already run Neo4j, FalkorDB, or Kuzu and want temporal memory on top of that infrastructure.

​
Coexistence

The two address different trust models. Teams that need extracted, time-aware memory from high-churn event streams pair Graphiti with a curated verified ledger: extraction handles recall breadth, while a Worlds world holds the assertions agents must be able to prove. Because Worlds is provider-agnostic, its graph remains queryable regardless of which extraction engine feeds other parts of the pipeline.