Key takeaways
- Temporal knowledge graph architecture that tracks how facts and relationships change over time, not just static embeddings
- Claims state-of-the-art agent memory performance, outperforming MemGPT/Letta on LongMemEval benchmarks by up to 18.5%
- Open-source Graphiti framework powers the core graph engine; commercial Zep Cloud handles orchestration and hosting
- Credit-based pricing starting at $25/month with enterprise BYOC options, but no self-hosted commercial option
FAQ
What is Zep?
Zep is a context engineering platform for AI agents that uses temporal knowledge graphs to provide long-term memory, GraphRAG, and context assembly. It automatically extracts entities, relationships, and facts from conversations and business data.
How does Zep differ from simple RAG or vector search?
Unlike vector-based retrieval, Zep builds a knowledge graph that tracks entity relationships and temporal changes. When facts change, old ones are invalidated. This enables relationship-aware queries and historical reasoning that vector search cannot provide.
Is Zep open source?
Zep's core graph engine, Graphiti, is open source (Apache 2.0) and can be self-hosted with Neo4j. The full Zep platform with managed memory, retrieval orchestration, and enterprise features is a commercial cloud service.
How does Zep pricing work?
Zep uses credit-based pricing at $25/month for 20,000 credits (1 credit per episode/message). A free tier offers 1,000 credits/month. Enterprise plans include managed, BYOK, BYOM, and BYOC deployment options.
Executive Summary
Zep is a context engineering and agent memory platform that uses temporal knowledge graphs to give AI agents persistent, relationship-aware memory. Unlike vector-based RAG approaches, Zep automatically extracts entities, relationships, and facts from conversations and business data, maintaining a graph that evolves over time and invalidates stale information. The company published a peer-reviewed paper demonstrating state-of-the-art performance on agent memory benchmarks.
| Attribute | Value |
|---|---|
| Company | Zep AI |
| Founded | 2023 |
| Open Source | Graphiti (Apache 2.0) |
| GitHub Stars | ~4.1K (getzep/zep), ~3K+ (getzep/graphiti) |
| Headquarters | USA |
How It Works
Zep operates in three stages: Ingest, Graph, and Assemble.
-
Ingest — Chat messages, JSON business data, and documents are sent to Zep as "episodes." The system automatically extracts entities, relationships, and facts using LLMs.
-
Graph — Extracted data feeds into a temporal knowledge graph. When facts change (e.g., a user moves cities), old edges are invalidated with timestamps rather than deleted, preserving historical context.
-
Assemble — When an agent needs context, Zep retrieves relevant subgraphs using semantic, keyword, and graph-based search, then formats results for token-efficient LLM consumption.
The underlying engine is Graphiti, Zep's open-source temporal knowledge graph framework. Graphiti supports Neo4j as its graph store and works with OpenAI, Azure OpenAI, Google Gemini, and Anthropic models. It also offers an MCP server for integration with Claude, Cursor, and other MCP-compatible clients.
Benchmark Performance
Zep claims an aggregate accuracy improvement of up to 18.5% over baseline on the LongMemEval benchmark, and outperforms MemGPT (Letta AI) on the Deep Memory Retrieval metric. The benchmark uses conversations averaging ~115,000 tokens.
Pricing
| Tier | Cost | Credits | Rate Limit |
|---|---|---|---|
| Free | $0/month | 1,000/month | Limited |
| Flex | $25/month | 20,000/month | 600 req/min |
| Enterprise | Custom | Custom | Guaranteed |
Credits map 1:1 to episodes (messages/data objects). Episodes larger than 350 bytes consume additional credits. Enterprise deployments include Managed, BYOK, BYOM, and BYOC options with SOC 2 Type II and HIPAA BAA compliance.
Strengths
- Temporal reasoning — Unlike static vector stores, the graph tracks when facts change and maintains history, enabling "what did we know when" queries
- Published research — Peer-reviewed arXiv paper with reproducible benchmarks gives credibility to performance claims
- Open-source core — Graphiti is Apache 2.0 licensed, usable independently with Neo4j for self-hosted setups
- Domain customization — Custom entity types and relationship models adapt to specific business domains
- Framework integrations — Works with LangChain, LlamaIndex, Vercel AI SDK, and offers MCP server support
Cautions
- Cloud-first model — The full Zep platform is cloud-only; self-hosting is limited to the Graphiti OSS layer without the orchestration features
- LLM dependency — Graph construction requires LLM calls for entity/relationship extraction, adding latency and cost on top of Zep's credit pricing
- Early performance issues — Reddit users reported initial performance problems in production, though Zep documented a 30x scaling improvement
- Credit-based pricing opacity — Large episodes consuming multiple credits makes cost prediction harder for high-volume use cases
- Neo4j dependency — Graphiti requires Neo4j, which adds infrastructure complexity for self-hosted deployments
Competitive Positioning
| Feature | Zep | Mem0 | Letta (MemGPT) | LangMem |
|---|---|---|---|---|
| Architecture | Temporal knowledge graph | Vector + graph hybrid | Agent-managed memory | LangChain-native |
| Open Source | Graphiti (Apache 2.0) | Yes (Apache 2.0) | Yes (Apache 2.0) | Yes |
| Self-Hosted | Graphiti only | Full | Full | Full |
| Temporal Reasoning | ✅ Native | Limited | Limited | No |
| Benchmark Claims | SOTA on LongMemEval | N/A | DMR benchmark creator | N/A |
| Managed Cloud | ✅ | ✅ | ✅ | Via LangSmith |
| Enterprise Options | BYOC, BYOK, BYOM | Enterprise tier | Enterprise tier | LangChain ecosystem |
Bottom Line
Zep represents the most technically ambitious approach to agent memory on the market. The temporal knowledge graph architecture is a genuine differentiator — most competitors rely on vector similarity search, which loses relationship context and has no concept of time. The published benchmarks are compelling, and the open-source Graphiti framework means teams can evaluate the core technology without vendor commitment.
The main tension is between Zep's strong open-source story (Graphiti) and its cloud-first commercial model. Teams wanting full self-hosted Zep — not just the graph engine — don't have that option. For production use, you're committing to Zep Cloud with credit-based pricing that can be hard to predict at scale. If your use case genuinely needs temporal reasoning about changing facts and relationships, Zep is the clear leader. If you just need "remember what the user said last week," simpler solutions like Mem0 may suffice.