Graphiti is Zep’s open-source framework for building real-time knowledge graphs for AI agents. Instead of treating memory as a flat transcript or replacing an old profile field with a new one, it models entities, relationships, provenance, and time. Episodes can arrive as messages, text, or JSON; Graphiti extracts facts and records both when a fact was valid in the real world and when the system learned it. This makes changing statements such as employment, account status, or project ownership queryable without erasing history. Compared with Mem0, it emphasizes temporal relationships and graph retrieval; compared with LangGraph, it is primarily a memory data layer rather than an execution orchestrator.
Quick Verdict
Graphiti is compelling when an agent must reason over evolving facts, connected entities, and point-in-time context. The official repository was active in 2026, and v0.29.2 was released on June 8, 2026. It is nevertheless an engineering framework, not a turnkey document chatbot. Teams own the graph database, model and embedding calls, indexes, backups, isolation, and correction workflow. For ordinary document Q&A, RAGFlow may be simpler and more directly useful.
Best For
It best serves AI engineers building customer-memory systems, investigation agents, changing entity profiles, or persistent assistants spanning many sessions. Platform teams should be comfortable operating a supported graph backend and defining tenant boundaries. It is a poor fit for no-code users, static FAQs, or projects where the extra extraction cost and graph operations cannot be justified by temporal queries.
Key Features
- Bitemporal facts separate real-world validity from system observation time, preserving invalidated relationships for historical queries.
- Incremental episodes add new messages, text, or JSON without rebuilding the entire graph.
- Hybrid search combines semantic vectors, BM25-style text retrieval, and graph traversal with filters.
- Entity and edge extraction uses LLMs and can be constrained with custom Pydantic entity types.
- Provenance links graph facts back to source episodes for explanation and correction.
- Backends and MCP include supported graph integrations and a separate MCP server; official release notes mark Kuzu deprecated.
Use Cases
- Track customers, employment, ownership, dependencies, or incidents whose state changes over time.
- Reconstruct what was valid or known at a particular point for review and investigation.
- Add durable context after each agent interaction instead of repeatedly sending a full transcript.
- Expose controlled graph-memory tools through MCP, with read and write permissions separated.
Pricing
| Option | Cost view | Responsibility |
|---|---|---|
| Graphiti library | Free under Apache-2.0 | Bring graph storage, LLMs, embeddings, monitoring, and operations |
| Graphiti MCP server | Open source with the project | Restrict clients, tools, credentials, and network exposure |
| Zep hosted product | Check Zep’s current terms | Do not assume hosted features or pricing are identical to Graphiti |
Pros
- Time is part of the fact model rather than an incidental metadata field.
- Incremental updates, source provenance, and hybrid retrieval form a coherent agent-memory layer.
- Apache-2.0 licensing, custom schemas, and multiple backends provide substantial engineering control.
- Recent releases and fixes provide a clear maintenance signal as of the review date.
Cons
- LLM extraction is fallible; important facts need validation, deduplication, and conflict handling.
- Graph storage, full-text indexing, model calls, backups, and migrations create a larger operating surface.
- An MCP server with mutation tools expands the trust boundary and should not be exposed broadly.
- Static document retrieval may gain little from the added complexity.
Alternatives
| Tool | Better for | Difference |
|---|---|---|
| Mem0 | Adding user-level long-term memory quickly | Simpler memory abstraction; Graphiti emphasizes temporal relationships |
| LangGraph | Stateful agent execution and checkpoints | Orchestrates flow state; Graphiti stores queryable long-term facts |
| RAGFlow | Parsed documents and citation-oriented RAG | Document-centric rather than changing entity relationships |
| LangChain | Composing models, retrievers, and tools | A general framework that can integrate Graphiti rather than directly replace it |
FAQ
Is Graphiti a graph database?
No. It is an application framework that builds and searches temporal knowledge graphs on a configured backend.
Does it replace conventional RAG?
Not universally. Temporal relationships are its strength; static document Q&A is often easier with a dedicated RAG system.
Why keep two kinds of time?
A fact’s real-world effective time can differ from when the system receives it. Separating them supports late events and historical reconstruction.
Can it run with local or compatible models?
Compatible configurations can be evaluated, but structured-output behavior, extraction quality, and performance must be tested for each model.
What permissions matter for MCP deployment?
Bind it to a controlled network, authenticate clients, separate reads from mutations, rotate secrets, and audit deletion and bulk-write operations.
Bottom Line
Graphiti turns agent memory into sourced, connected facts with explicit validity periods. That is materially better than transcript summaries when a business genuinely needs changing entities and historical reconstruction. Production success still depends on tenant isolation, graph operations, model permissions, correction, and retention. If those needs are absent, start with a lighter memory or RAG layer.