> ## 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.

# Supermemory

> Worlds vs Supermemory: extracted memory against curated graphs.

## At a glance

|                 | Worlds                                           | [Supermemory](https://supermemory.ai) |
| :-------------- | :----------------------------------------------- | :------------------------------------ |
| Data model      | RDF triples in an append-only ledger             | Extracted entities and documents      |
| How facts enter | Curated import (quads or text)                   | Automatic extraction from sources     |
| Retrieval       | Hybrid: vector, full-text, RDF filters           | Search, memory graph, profiles        |
| Provenance      | Fact-level, chronological, verifiable via SPARQL | Partial, derived from sources         |
| Deployment      | Hosted beta or embedded adapters                 | Hosted or self-hosted                 |

## The philosophical difference

[Supermemory](https://supermemory.ai) builds memory by extracting and
structuring information from your connected sources, then serving it back
through search, a knowledge graph, and entity profiles. Extraction does the
work: you connect a source and the memory layer derives the structure.

Worlds takes the opposite stance. Memory is curated, not accumulated. You decide
which facts matter, assert them as RDF triples, and the engine keeps them in an
append-only ledger. Nothing is inferred; everything is stated.

The tradeoff is effort against control. Extraction-based memory is cheap to
start and useful for loosely-structured recall. Curated graphs cost more to
maintain and return exactly what you stored, with a path back to every source.

## When Worlds fits

* The fact must be true, not merely similar: compliance, safety, operations.
* You need to resolve contradictions at the data layer instead of leaving the
  model to guess between conflicting chunks.
* You want a deterministic [query language](/worlds/query) (SPARQL) over your
  knowledge, not just retrieval.
* You need a chronological record of when facts changed and who asserted them.

## When Worlds does not fit

* You want zero-effort memory from messy, unstructured sources with no curation
  budget.
* Your use case is loose recall ("what did we talk about last week") where an
  extracted summary is enough.
* You are prototyping and have no graph vocabulary in mind yet.

## Coexistence

The [MemSDK](/projects/memsdk) defines a common memory interface. You can switch
which backend serves a client without changing the calling code, which lets you
prototype against an extraction-based store and move to Worlds when the facts
need to be authoritative.