Zep ships a temporal conversational memory. REM Labs is the full continuity layer — nine consolidation strategies, federation, reactivity, and retrieval that beats theirs cold and warm. 94.6% on LongMemEval (vs Zep's 63.8%) under the byte-exact upstream GPT-4o judge, 78ms cold retrieval on 1M-memory corpora, and full self-host in one command. Here's the breakdown.
Zep is a well-engineered product. So is REM — and the numbers show it.
Zep stores and retrieves facts well. REM runs nine consolidation strategies over those facts every night — and federates them across agents, with webhook reactivity baked in. Memory should evolve, not just sit there.
Zep is strong on two pillars (persist + federate, within a single agent). REM spans all four: Persist · Evolve · Federate · React. Consolidation, cross-agent namespaces, and webhook reactivity are the three pillars Zep doesn't ship today.
Zep is a commercial SaaS with an OSS core (Graphiti). Numbers below are from their docs, the published Graphiti paper, and public LongMemEval results.
| Dimension | REM Labs | Zep |
|---|---|---|
| Category | Continuity layer for intelligence | Conversational memory with temporal graph |
| LongMemEval (500q) | 94.6% · byte-exact upstream GPT-4o judge | 63.8% (Graphiti paper, 2024) |
| Consolidation strategies | 9 (Dream Engine) | 1 (fact extraction + graph update) |
| Cold-start retrieval p50 | 78ms cold / 42ms warm (edge-cached hot index) | ~200ms |
| Model-agnostic | Yes — every LLM vendor + local | Yes — OpenAI, Anthropic, local |
| Self-hostable | Yes — Docker + K8s + bare metal, one command, ~90s, unlimited everything | Partial — Graphiti OSS + Zep community edition |
| Open source | Apache 2.0 core — SDKs + self-host + extractors | Partial — Graphiti OSS, Zep Cloud closed |
| GDPR / forget API | Yes — per-memory + audit | Yes — session + user delete |
| Federation across agents | Yes — shared namespaces + A2A | Partial — groups, but no cross-agent consolidation |
| Webhooks / reactivity | Yes — memory / dream / contradiction events | No native; poll the API |
| MCP / A2A protocol | Yes | No native MCP endpoint |
| Multi-agent / hive | Yes — DreamHive coordinated consolidation | Partial — user groups share graph facts |
| Pricing start | Free (unlimited memories, 500 dreams/mo) → $19 Pro | Free tier; Cloud paid plans from ~$20-50/mo tier |
| Temporal graph | Yes — Temporal Merge + Contradiction + Cross-Memory Association with time-travel queries | Yes — Graphiti |
LONGMEMEVAL METHODOLOGY · /benchmarks
The two dimensions Zep markets hardest — and REM's actual numbers on each.
REM hits 78ms cold / 42ms warm on 1M-memory corpora via an edge-cached hot index. Zep's best public number is ~200ms. REM leads on one-shot lookups, warm recall, and the long tail alike.
REM runs Temporal Merge, Cross-Memory Association, and Contradiction Resolution as three of its nine consolidation strategies — giving you time-scoped facts, auto-resolved conflicts, and synthesis across 200+ memories overnight. Graphiti does the first; REM does all four.
Zep as your low-latency retrieval tier, REM as the consolidation layer that runs on Zep's graph overnight. Graphiti's fact edges → REM's 9 strategies → updated fact edges pushed back. We publish the integration recipe on request.
Bring your retrieval layer if you want — or let REM handle the whole stack. Free tier, no credit card.