Zep vs MemMesh: A Temporal Graph vs. a Brain That Forecasts

Of everyone in the memory field, Zep is the one whose architecture we most respect. Its open-source engine, Graphiti (Apache-2.0), builds a bi-temporal knowledge graph — it tracks not just what's true, but when a fact was true in the world versus when the system learned it. That's a rigorous, rare distinction, and it makes point-in-time queries correct in a way most memory layers can't manage. If you're choosing a memory system on the strength of its graph, Zep belongs at the top of your list.

So this comparison isn't about beating Zep at its own game. It's about two axes Zep doesn't play on. Everything here is from Zep's and Graphiti's public docs and the Zep paper as of July 2026.

What Zep does well — credit where it's due

• Bi-temporal graph — facts carry validity windows; when a fact changes, the old edge is invalidated, not deleted, so history is preserved.
• Strong provenance — every derived fact traces back to the source episode it came from. This is an area where Zep is genuinely excellent.
• Hybrid retrieval — semantic embeddings + BM25 keyword + graph traversal, with sub-200ms recall at large scale.
• Mature enterprise governance — SOC 2 Type II, HIPAA BAA, audit logs, retention, and BYOC deployment.

Note what's not on our differentiator list: provenance. Zep has it, and we won't pretend otherwise. Where we'd distinguish ourselves isn't whether provenance exists, but what a brain does with it.

Difference 1: prediction

Zep recalls, invalidates stale facts, and surfaces patterns in the graph it already holds. What it doesn't do is forecast a future event with calibrated confidence. A MemMesh brain declares a target and predicts it from a subject's history, scores that confidence against what actually happens, and recalibrates. Zep tells you what's true now, with an audit trail; a brain also tells you what's likely next, with a number you can act on.

Difference 2: installable, publishable brains

Zep's memory is scoped per tenant — powerful within a customer, but not a shareable asset. There's no way to package a domain-specialized memory and let another team install or buy it. MemMesh makes a brain a first-class unit you publish over the Mesh Router, so any model can install it with per-brain provenance. Expertise becomes distributable, not just queryable.

Difference 3: on-device posture

Zep Cloud is a hosted SaaS, and even self-hosted Graphiti typically calls external models for extraction and needs a managed graph database. A MemMesh brain runs on-device and private by default. For teams that need data to stay put, that's a structural difference, not a config flag.

When to pick Zep

If your problem is fundamentally a temporal-graph problem — you need rigorous point-in-time correctness, deep relationship traversal, and mature enterprise compliance, and you're comfortable on a hosted or BYOC deployment — Zep is an excellent choice, and Graphiti's open core means no lock-in at the engine layer. One caveat to check: the fully self-hostable 'Community Edition' of the product has been reported deprecated, so 'self-host the whole product for free' may not be on the table.

Zep answers 'what is true, and when?' with an audit trail. A brain also answers 'what happens next?' — and can be handed to any model that needs it.

Choose Zep for the temporal graph. Choose MemMesh when you also need calibrated foresight, an on-device footprint, or a brain you can publish and install. We'll keep this honest as both evolve.