What Worlds provides
- Graph storage: RDF triples in an append-only, chronological fact ledger.
- Hybrid search: vector, full-text, and RDF filters fused by reciprocal rank fusion.
- SPARQL queries: declarative, deterministic graph traversal.
- Verifiable provenance: every fact carries a traceable path to its source.
- Provider-agnostic: you connect your own LLMs and embedding providers.
- Edge-ready adapters: LibSQL/Turso and Postgres backends.
What Worlds does not do
Worlds is deliberately narrow. It does not:- Host an LLM. You bring the model. Worlds stores and retrieves facts; the model decides what to ask.
- Host an embedding service. Embeddings come from the embedding provider you configure.
- Extract memory automatically. Memory in Worlds is curated. Facts enter a world deliberately, through import and patches, rather than being inferred from transcripts.
- Run agents. Worlds is not an agent runtime. It is a context source that agents query through tool calls.
- Replace your primary database. Worlds is a context engine, not a general operational store. Adapters persist to your existing substrate.
- Equate vector similarity with relevance. Similarity != relevance. Vectors are only one retrieval signal in Worlds. The graph structure and SPARQL filters ensure agents retrieve exact, verifiable facts rather than loose semantic guesses.
How the alternatives line up
Polygres adds graph-style retrieval over Postgres records rather than RDF
triples. Turso is the substrate underneath Worlds’
worlds-libsql adapter
rather than a direct alternative.
The remaining alternatives cluster by model. Similarity layers such as LangMem,
MemSearch, and MemSync recall by embedding. Graph-shaped stores such as
MemMachine combine a knowledge graph with SQL. Governed services such as Twilio
Memory Store, MemClaw, and Maindex organize structure and access. MemSpan trades
structure for portability. None of them exposes a queryable, append-only fact
ledger with deterministic SPARQL verification.
When to choose Worlds
Choose Worlds when conformance and interoperability are non-negotiable: agents that rely on W3C open standards (RDF, SPARQL), portable memory layers, and deterministic graph verification across any database or model provider. The append-only ledger and SPARQL query engine keep agent context interoperable and inspectable across database and model providers. Real-world problems Worlds solves:- Your AI coding assistant forgets your codebase between sessions. Chat logs truncate and vector recall drifts from the facts. Store facts about your APIs, conventions, and past decisions in a world so the next session resumes from verified graph state instead of a truncated conversation log.
- Your multi-step agent drifts from the facts mid-task. Long task chains accumulate errors when each step reasons from fuzzy similarity. SPARQL pattern matching confirms that entities and relationships exist before the agent acts on them, keeping multi-hop reasoning grounded in facts.
- You cannot answer “where did this fact come from?” In strict environments, every assertion needs a traceable source. The append-only ledger records quad-level provenance, so you can audit where each fact came from and when it entered the world.
- You are locked into one vector store or model provider. Worlds is provider-agnostic: you bring your own LLMs and embedding providers, and adapters persist to LibSQL/Turso or Postgres, so context stays portable.