Compare

Mnemix vs Zep

The research-cited choice. Bi-temporal knowledge graph (Graphiti), SOC 2, BYOC. Voice is a community demo only.

Motion available.Caller identity wedge: phone ring to informed responseThe phone rings, enrichment fans out through Twilio Lookup, Trestle, and Baylio intent, then the agent's first response is already informed.Phone ringsEnrichment fans out in parallelAgent's first response is already informed+1 555 123 4567Twilio LookupTrestleBaylio intentAgentHi Mike, I see this is about yourfleet account...Real-world identity, intent, and history — joined to your agent's memory — at Cloudflare edge latency.

Animation paused for reduced motion

60-second verdict

Zep's bi-temporal graph is the strongest research-cited memory architecture for chat agents and Graphiti is genuinely impressive engineering. But voice is a community demo, not a first-class integration, and the Mem0 paper's independent LOCOMO result (63.8%) sits well below Zep's self-reported 71.2%. Choose Mnemix if you're building voice.

Side-by-side

DimensionZepMnemix
GitHub stars25,463 (graphiti) + 4,495 (zep)n/a
Bi-temporal✅ (4 timestamps in graphiti_core/edges.py)✅ (session-scoped)
Voice integrations🟡 community demo only✅ Twilio, Vapi, Retell, Bland
Caller-ID enrichment✅ Twilio Lookup + Trestle + Baylio
Edge runtime❌ Python on AWS✅ Cloudflare Workers
Claimed P95 retrieval<200msdesigned for sub-300ms voice recall
LongMemEval (self)71.2%Coming May 2026
LongMemEval (independent)63.8% (Mem0 paper)Coming May 2026
Starter price$25 to $125 (5x)Hobby $0; Starter+ contact sales

When you'd pick Zep

When you'd pick Zep: you're building a chat agent that needs deep graph queries over conversation history, you have AWS BYOC requirements, your latency budget is generous (~200ms is fine), and SOC 2 Type 2 today matters more than voice integrations.

When you'd pick Mnemix

When you'd pick Mnemix: voice is the channel. You need the same bi-temporal model Zep has, but at Cloudflare edge latency, with caller identity bundled in, and integration kits maintained for Vapi, Retell, Bland, and LiveKit on day one.

FAQ

Zep's bi-temporal graph is research-grade. What does Mnemix have?
Mnemix also stores four timestamps per fact (valid_from, valid_to, observed_at, ingested_at) — the same bi-temporal model. The difference is operational: Zep runs Python on AWS, Mnemix runs on Cloudflare Workers, and Mnemix bundles caller-ID resolution. For a chat-only agent with deep graph queries, Zep is excellent. For voice, the latency budget pushes you to the edge.
Will Zep ship first-class voice eventually?
They might — and if they do, the comparison shifts. Today, Zep's voice support is community examples and demos rather than maintained integration kits with Vapi/Retell/Bland. Mnemix shipped with voice as the only target.
What about Zep's independent LOCOMO score?
Zep self-reports 71.2% on LongMemEval; the Mem0 paper's independent measurement put them at 63.8%. We're not litigating which is correct — but it's why Mnemix is publishing methodology + harness alongside our v1.0 score in May 2026, so independent replication is one command away.
Can I run both Zep and Mnemix together?
Yes. We've seen teams use Zep for chat history + graph queries and Mnemix for the voice channel + caller identity. The two share well at the application layer; we provide a /v1/import endpoint if you want to consolidate later.

Ship Mnemix in 5 minutes

Free Hobby tier — 50 sessions, 1,000 memory ops, 100 Twilio Lookups.

Get a key — free

Last updated: .