Neotoma vs Zep
Zep is a retrieval memory system that combines vector search with knowledge graphs for AI assistant context. Neotoma is a deterministic state layer that provides versioned, schema-bound, auditable entity state. Zep optimizes for retrieval relevance; Neotoma optimizes for state integrity.
Both Zep and Neotoma provide persistent memory for AI agents. They take fundamentally different approaches to the memory problem.
How does Neotoma compare to Zep?
Zep
Zep combines vector similarity search with a knowledge graph extracted from conversations. It auto-summarizes sessions, extracts entities, and builds a temporal knowledge graph. Memory is optimized for retrieving relevant context with business-data enrichment.
Neotoma
Neotoma stores append-only observations and composes them into entity snapshots via deterministic reducers. Every state change is versioned with full provenance. Memory is optimized for knowing exactly what was true at any point in time.
Guarantee comparison
When to use which
Use Zep when
You need enriched retrieval with automatic knowledge graph extraction from conversations. Your primary concern is injecting the most relevant context into prompts with minimal configuration.
Use Neotoma when
You need formal state integrity guarantees: deterministic reconstruction, multi-writer conflict resolution, schema validation, auditable provenance, and temporal state queries.
Common questions
Does Zep provide deterministic state reconstruction?
No. Zep's knowledge graph is built via extraction and summarization, which are non-deterministic processes. The same inputs may produce different graph states across runs.
Can I use Zep and Neotoma together?
Yes. Zep handles retrieval-augmented context for prompts while Neotoma handles structured state integrity for entities that need formal guarantees.