At a glance
The philosophical difference
Zep keeps long-term memory for agents by capturing conversation history and extracting entities and facts over time. Retrieval returns what is similar to the current context, with temporal awareness baked into the memory model. Worlds stores facts you assert as RDF triples in an append-only ledger. The engine does not watch conversations and extract from them. You curate, the graph holds, and SPARQL verifies. Temporal memory is valuable for chat continuity: the agent remembers what happened earlier in the session. It is a weaker foundation for acting on facts, because the recalled item is a derived memory, not a fact with a verifiable source.When Worlds fits
- The agent’s actions depend on facts that must be exact and checkable.
- You want one source of truth shared across sessions, agents, and models.
- Contradicting facts need resolution at the data layer, not by the model.
When Worlds does not fit
- You need rich conversation recall with minimal setup.
- Factual exactness is less important than remembering context.
- You have no curation budget for maintaining a graph vocabulary.