> ## Documentation Index
> Fetch the complete documentation index at: https://docs.wazoo.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# Cognee

> Worlds vs Cognee: curated RDF fact ledger vs ECL-extracted graph and vector memory.

## At a glance

|              | Worlds                                                | [Cognee](https://github.com/topoteretes/cognee)            |
| :----------- | :---------------------------------------------------- | :--------------------------------------------------------- |
| Primary role | Authoritative RDF fact ledger & SPARQL context engine | Self-hosted agent-memory platform over graphs plus vectors |
| Data model   | W3C RDF triples in an append-only graph ledger        | Extracted entity-relation graph with embeddings (ECL)      |
| Retrieval    | SPARQL graph queries, full-text, & vector fusion      | Graph traversal combined with similarity search            |
| Ingestion    | Curated imports and patches                           | Automated extract-cognify-load pipeline over many sources  |
| Deployment   | Edge-ready adapters (LibSQL, Postgres) or Cloud       | Python SDK, Docker, or Cognee Cloud; pluggable backends    |

## Overlap and shared capabilities

Both Worlds and [Cognee](https://cognee.ai) give agents persistent context that
goes past a single session. Both combine graph structure with vector signals so
recall can follow relationships between entities instead of returning only
similar text.

The difference is in what the graph is and who authors it:

* Cognee runs an automated pipeline: ingest data in any format, extract entities
  and relationships, cognify them into a queryable graph with embeddings, and
  serve recall through `remember` / `recall` operations. The graph is derived
  output; its accuracy tracks the extraction models.
* Worlds keeps the graph authoritative. Facts enter as W3C RDF through import
  and patches you control, land in an append-only ledger with per-quad
  provenance, and verify through SPARQL rather than similarity.

Cognee's automation is the feature when volume is high and curation is not an
option. When assertions must be provable, for example in audit, compliance, or
multi-team settings, the extraction step becomes the liability: a misread at
ingest time propagates into every downstream answer.

## When Worlds fits

* You need deterministic SPARQL verification of entities and relationships
  before an agent acts on them.
* You require provenance for every fact: where it came from and when it entered
  the world.
* You need standards-based portability across LibSQL/Turso and Postgres backends
  with your own model and embedding providers.

## When Cognee fits

* You want hands-off memory: point the pipeline at documents, chats, or APIs and
  let extraction build the graph.
* You need flexible deployment across local runs, self-hosted Docker, air-gapped
  environments, or managed cloud.
* Your agents benefit from ontology-guided extraction over heterogeneous sources
  where manual curation cannot keep up.

## Coexistence

The two occupy different points on the curation spectrum. A common pattern is
extraction-first ingestion with verified storage downstream: Cognee's pipeline
processes high-volume raw sources, while a [Worlds](/projects/worlds) world
holds the curated assertions that agents must verify through SPARQL before
acting. Worlds' provider-agnostic design keeps the verified layer independent of
whichever extraction engine feeds it.