Zep and Graphiti
Graphiti (open source, Apache-2.0, 32,885 stars on 2026-11-17) is a Python framework for temporal context graphs. Zep is the commercial “unified context layer for enterprises” built on it — managed or in your own cloud (BYOC), with its own graph store (“Konig”), ingestion, governance and serving. Graphiti is self-hosted only and you bring your own graph database. Paper: arXiv 2501.13956.
Design
- Facts are edges with validity windows; changed facts are invalidated, not deleted, so history is kept.
- Every entity/fact traces back to source episodes (provenance).
- Ontology via Pydantic models, learned, or both; incremental updates, no batch recompute.
- Retrieval: embeddings + BM25 + graph traversal.
- Backends: Neo4j 5.26, FalkorDB (incl. embedded Lite), Amazon Neptune + OpenSearch. See Neo4j.
- Reference MCP and REST servers exist but have no auth or prompt-injection protection — not production.
- Telemetry (PostHog) on by default; disable with
GRAPHITI_TELEMETRY_ENABLED=false.
Repository snapshot (checked 2026-11-17)
| Item | Detail |
|---|---|
| Repo | getzep/graphiti, Python, Apache-2.0 |
| Created | 2024-08-08 |
| Activity | last push 2026-11-15; latest release v0.31.2 (2026-10-20, a certifi update) |
| Size | 32,885 stars, 3,383 forks |
| Package | graphiti-core on PyPI, with extras falkordb, falkordblite (Python 3.12+), neptune, anthropic, groq, google-genai |
| Docs | https://help.getzep.com/graphiti |
What a context graph holds
| Component | Stores |
|---|---|
| Entities (nodes) | People, products, policies and concepts, with summaries that evolve |
| Facts (edges) | Entity, relationship, entity triplets with a validity window |
| Episodes | The raw ingested data; every derived fact traces back to one |
| Custom types | Developer-defined entity and edge types as Pydantic models |
Requirements and setup
- Python 3.10 or higher.
- A graph database: Neo4j 5.26, FalkorDB 1.1.2, or Amazon Neptune plus an Amazon OpenSearch Serverless collection for full-text search.
- An LLM and an embedder. OpenAI is the default and needs
OPENAI_API_KEY. Anthropic, Gemini and Groq are supported, and Azure OpenAI works through Azure’s v1 API. Any OpenAI-compatible endpoint (DeepSeek, Together, OpenRouter, Ollama, vLLM, llama.cpp, LM Studio) works throughOpenAIGenericClient. - Structured output is required for extraction and deduplication. The README warns that small or local models often emit JSON that fails the schema, and offers a
structured_output_modeofjson_schema(default) orjson_object. - Concurrency is capped by
SEMAPHORE_LIMIT, default 10, to avoid provider 429 errors. Raise it for speed, lower it if rate-limited. - Docker Compose starts Neo4j, or FalkorDB with
--profile falkordb.
Features worth knowing
identity_propertiesmatch entities by chosen fields (for example a Person by email) before node resolution; supported on Neo4j and FalkorDB.- Prompt builders can be overridden per instance, and an opt-in
llm_runtimeroutes individual prompts to different models and providers. Response schemas stay fixed. - Reference servers:
mcp_server(episode, entity and search tools over MCP) andserver(FastAPI REST). The README says they are not production deployments. For production, usegraphiti-corein your own service, which must supply authentication, tenant isolation and input controls, or use Zep.
Zep versus Graphiti (from the README)
| Aspect | Zep | Graphiti |
|---|---|---|
| What it is | Enterprise context layer, managed or in your cloud (BYOC) | Open-source framework, self-hosted only |
| Graph database | Konig, Zep’s own service | Bring your own: Neo4j, FalkorDB or Neptune |
| Users, threads, ingestion | Built in | Build your own |
| Retrieval | Pre-configured; the README claims sub-200ms at scale | Custom; depends on your setup |
| Tooling | Dashboard, CLI, SDKs for Python, TypeScript and Go | Build your own |
Graphiti versus GraphRAG (the vendor’s comparison)
The README contrasts Graphiti’s continuous incremental updates, explicit bi-temporal tracking with automatic fact invalidation, and hybrid retrieval with GraphRAG’s batch processing, community summaries and sequential LLM summarization. It rates Graphiti as typically sub-second and GraphRAG as seconds to tens of seconds. These are the authors’ claims, not measured here. See graphrag.
Benchmarks
Published scores for Graphiti and Zep differ by source, test harness and judge model, so no single figure is quoted here. The comparison across memory systems is in agent-memory-benchmarks.
Take
Its distinguishing feature is facts with validity windows, so it can answer “what was true when”. The cost is heavier operation: you run a graph database, an LLM and an embedder, and build the service around it. Compare mem0, graphrag, cognee; landscape in agent-memory-systems.