Both products take memory seriously. But RetainDB handles three layers — user memory, context assembly, and knowledge base ingestion (Notion, PDFs, Confluence, YouTube) — not just memory. Zep goes deep on entity graphs and temporal reasoning. RetainDB goes deep on recall quality, knowledge ingestion, and production simplicity.
Zep's temporal knowledge graph is genuinely impressive engineering. RetainDB scores 31 points higher on preference recall. Pick based on what your users will actually feel — not the architecture you find more interesting.
Preference recall measures whether your agent correctly remembers what a user told it. RetainDB: 88%. Zep: 56.7%. That gap means RetainDB correctly recalls a preference on roughly 1 in 3 queries that Zep misses. Both scores from the same LongMemEval task set — retaindb.com/benchmark.
Zep's graph traversal is optimised for entity-relationship reasoning. For preference memory and conversational continuity, vector + BM25 + reranking achieves higher recall — the BM25 branch catches phrasing that embedding distance buries.
Run npx retaindb-wizard, it detects your framework, generates integration code. With Zep, graph architecture — node types, edge definitions, traversal strategy — has to be designed before you write a single memory.
Zep's graph is built for entity memory extracted from conversations. If you also need your agents to know your product documentation, help center, research papers, or Notion workspace, that's a separate integration with Zep. RetainDB's 15+ built-in connectors (Notion, Confluence, PDF, YouTube, arXiv, Playwright, sitemaps) make knowledge and memory retrieval part of the same system — composed per query by type and scope.
Both scores come from the same LongMemEval preference recall task. RetainDB published methodology at retaindb.com/benchmark. Zep's score is from the same benchmark surface. Verify it yourself — the methodology is public.
Yes — optional graph traversal via include_graph with configurable depth. It's not the primary retrieval path, but entity relationship data is stored and queryable.
Graph traversal is optimised for entity-relationship reasoning, not preference string recall. 'I prefer dark mode' is better recalled by hybrid search over typed preference memories than by graph edge traversal.
88% preference recall on LongMemEval. Under 40ms retrieval. Most teams are in production in under 30 minutes — no infrastructure to manage.
Pages that keep the comparison moving deeper into the RetainDB memory and context cluster.