Home / Compare / Zep / Graphiti

Lians vs Zep / Graphiti.

Zep / Graphiti is the temporal knowledge graph, scored on the same five invariants we score ourselves. Credit where it's strong, receipts where it isn't.

Credit first

Where Zep / Graphiti is genuinely strong.

Zep is the strongest competitor on this eval, and honestly so: Graphiti extracts entities and edges with bitemporal validity stamps, invalidates contradicted edges, and supports temporal filtering. If your problem is a conversational knowledge graph, Zep is a serious tool.

Regulated-memory eval · last run 2026-07-04

The five invariants.

Regulated invariantLiansZep / Graphiti
Stale revision suppressedpasspartial
Point-in-time (as-of) recallpasspartial
Provable erasure (crypto-shred + cert)passpartial
Lookahead / backtest guardpassabsent
Audit-state snapshot at Tpasspartial
Score (pass = 1, partial = ½)5.0 / 52.0 / 5

Both columns executed live: Graphiti OSS in its default documented config. Its contradiction invalidation genuinely works, but default search still returns invalidated edges unless the caller filters them. Evidence in the methodology doc ↗.

Where the lanes diverge

Different jobs, different guarantees.

The gap is turnkey compliance: temporal filtering is not an as-of recall primitive, edge invalidation depends on LLM extraction quality, deletion has no proof artifact, and there is no lookahead guard or first-class snapshot API. Lians also enforces isolation in Postgres RLS below the app, and puts an admission gate in front of every write.

The honest recommendation

Which should you pick?

Choose Zep for graph-shaped conversational memory where extraction quality matters more than provability. Choose Lians when the temporal model must be deterministic, auditable, and erasable with a certificate.

A comparison doc exists: Lians vs Zep, in depth ↗. Lians ships a relationship graph on the same compliance spine, so graph queries carry over.

Bring the numbers to your compliance officer.

Get an API key →