getzep/graphiti

⭐ Help us reach more developers and grow the Graphiti community. Star this repo!

Graphiti is a framework for building and querying temporal context graphs for AI agents. Unlike static knowledge graphs, Graphiti's context graphs track how facts change over time, maintain provenance to source data, and support both prescribed and learned ontology — making them purpose-built for agents operating on evolving, real-world data.

Unlike traditional retrieval-augmented generation (RAG) methods, Graphiti continuously integrates user interactions, structured and unstructured enterprise data, and external information into a coherent, queryable graph. The framework supports incremental data updates, efficient retrieval, and precise historical queries without requiring complete graph recomputation, making it suitable for developing interactive, context-aware AI applications.

Use Graphiti to:

• Build temporal knowledge graphs that evolve with every interaction — tracking what's true now and what was true before.
• Represent rich, structured context instead of flat document chunks or raw event streams.
• Query across time, meaning, and relationships with hybrid retrieval (semantic + keyword + graph traversal).

A context graph is a temporal graph of entities, relationships, and facts — like "Kendra loves Adidas shoes (as of March 2026)." Unlike traditional knowledge graphs, each fact in a context graph has a validity window: when it became true, and when (if ever) it was superseded. Entities evolve over time with updated summaries. Everything traces back to episodes — the raw data that produced it.

What makes Graphiti unique is its ability to autonomously build context graphs from unstructured and structured data, handling changing relationships while preserving full temporal history.

A context graph contains:

Graphiti is the open-source framework for building temporal knowledge graphs at the core of Zep's context infrastructure. Zep manages context graphs at scale, providing governed, low-latency context retrieval and assembly for production deployments.

Under the hood, Zep is powered by a proprietary graph database — the Context Graph Engine — built for millions of context graphs with low-latency retrieval, so production deployments don't require a separate third-party graph database.

Read our paper: Zep: A Temporal Knowledge Graph Architecture.

Choose Zep if you want a turnkey, enterprise-grade platform with security, performance, and support baked in.

Choose Graphiti if you want a flexible OSS core and you're comfortable building/operating the surrounding system.

Traditional RAG approaches often rely on batch processing and static data summarization, making them inefficient for frequently changing data. Graphiti addresses these challenges by providing:

• Temporal Fact Management: Facts have validity windows. When information changes, old facts are invalidated — not deleted. Query what's true now, or what was true at any point in time.
• Episodes & Provenance: Every entity and relationship traces back to the episodes (raw data) that produced it. Full lineage from derived fact to source.
• Prescribed & Learned Ontology: Define entity and edge types upfront via Pydantic models (prescribed), or let structure emerge from your data (learned). Start simple, evolve as patterns appear.
• Incremental Graph Construction: New data integrates immediately without batch recomputation. The graph evolves in real-time as episodes are ingested.
• Hybrid Retrieval: Combines semantic embeddings, keyword (BM25), and graph traversal for low-latency, high-precision queries without reliance on LLM summarization.
• Scalability: Efficiently manages large datasets with parallel processing, pluggable graph backends, suitable for enterprise workloads.

Graphiti is specifically designed to address the challenges of dynamic and frequently updated datasets, making it particularly suitable for applications requiring real-time interaction and precise historical queries.

Requirements:

• Python 3.10 or higher
• Neo4j 5.26 / FalkorDB 1.1.2 / Amazon Neptune Database Cluster or Neptune Analytics Graph + Amazon OpenSearch Serverless collection (serves as the full text search backend) / Kuzu 0.11.2 (deprecated, see below)
• OpenAI API key (Graphiti defaults to OpenAI for LLM inference and embedding)

Optional:

• Google Gemini, Anthropic, or Groq API key (for alternative LLM providers)

docker run -p 6379:6379 -p 3000:3000 -it --rm falkordb/falkordb:latest

pip install graphiti-core

or

uv add graphiti-core

If you plan to use FalkorDB as your graph database backend, install with the FalkorDB extra:

pip install graphiti-core[falkordb]

# or with uv
uv add graphiti-core[falkordb]

# or embedded version (requires Python 3.12+)
pip install graphiti-core[falkordblite]
# or with uv
uv add graphiti-core[falkordblite]

If you plan to use Kuzu as your graph database backend, install with the Kuzu extra:

pip install graphiti-core[kuzu]

# or with uv
uv add graphiti-core[kuzu]

If you plan to use Amazon Neptune as your graph database backend, install with the Amazon Neptune extra:

pip install graphiti-core[neptune]

# or with uv
uv add graphiti-core[neptune]

# Install with Anthropic support
pip install graphiti-core[anthropic]

# Install with Groq support
pip install graphiti-core[groq]

# Install with Google Gemini support
pip install graphiti-core[google-genai]

# Install with multiple providers
pip install graphiti-core[anthropic,groq,google-genai]

# Install with FalkorDB and LLM providers
pip install graphiti-core[falkordb,anthropic,google-genai]

# Install with Amazon Neptune
pip install graphiti-core[neptune]

Graphiti's ingestion pipelines are designed for high concurrency. By default, concurrency is set low to avoid LLM Provider 429 Rate Limit Errors. If you find Graphiti slow, please increase concurrency as described below.

Concurrency controlled by the SEMAPHORE_LIMIT environment variable. By default, SEMAPHORE_LIMIT is set to 10 concurrent operations to help prevent 429 rate limit errors from your LLM provider. If you encounter such errors, try lowering this value.

If your LLM provider allows higher throughput, you can increase SEMAPHORE_LIMIT to boost episode ingestion performance.

For a complete working example, see the Quickstart Example in the examples directory. The quickstart demonstrates:

1. Connecting to a Neo4j, Amazon Neptune, FalkorDB, or Kuzu database
2. Initializing Graphiti indices and constraints
3. Adding episodes to the graph (both text and structured JSON)
4. Searching for relationships (edges) using hybrid search
5. Reranking search results using graph distance
6. Searching for nodes using predefined search recipes

The example is fully documented with clear explanations of each functionality and includes a comprehensive README with setup instructions and next steps.

You can use Docker Compose to quickly start the required services:

• Neo4j Docker:

  docker compose up

  This will start the Neo4j Docker service and related components.

• FalkorDB Docker:

  docker compose --profile falkordb up

  This will start the FalkorDB Docker service and related components.

The mcp_server directory contains a Model Context Protocol (MCP) server implementation for Graphiti. This server allows AI assistants to interact with Graphiti's context graph capabilities through the MCP protocol.

Key features of the MCP server include:

• Episode management (add, retrieve, delete)
• Entity management and relationship handling
• Semantic and hybrid search capabilities
• Group management for organizing related data
• Graph maintenance operations

The MCP server can be deployed using Docker with Neo4j, making it easy to integrate Graphiti into your AI assistant workflows.

For detailed setup instructions and usage examples, see the MCP server README.

The server directory contains an API service for interacting with the Graphiti API. It is built using FastAPI.

Please see the server README for more information.

In addition to the Neo4j and OpenAi-compatible credentials, Graphiti also has a few optional environment variables. If you are using one of our supported models, such as Anthropic or Voyage models, the necessary environment variables must be set.

Database names are configured directly in the driver constructors:

• Neo4j: Database name defaults to neo4j (hardcoded in Neo4jDriver)
• FalkorDB: Database name defaults to default_db (hardcoded in FalkorDriver)

As of v0.17.0, if you need to customize your database configuration, you can instantiate a database driver and pass it to the Graphiti constructor using the graph_driver parameter.

from graphiti_core import Graphiti
from graphiti_core.driver.neo4j_driver import Neo4jDriver

# Create a Neo4j driver with custom database name
driver = Neo4jDriver(
    uri="bolt://localhost:7687",
    user="neo4j",
    password="password",
    database="my_custom_database"  # Custom database name
)

# Pass the driver to Graphiti
graphiti = Graphiti(graph_driver=driver)

from graphiti_core import Graphiti
from graphiti_core.driver.falkordb_driver import FalkorDriver

# Create a FalkorDB driver with custom database name
driver = FalkorDriver(
    host="localhost",
    port=6379,
    username="falkor_user",  # Optional
    password="falkor_password",  # Optional
    database="my_custom_graph"  # Custom database name
)

# Or use embedded FalkorDB Lite (requires Python 3.12+)
# from redislite.async_falkordb_client import AsyncFalkorDB
# falkordb_client = AsyncFalkorDB(dbfilename='/path/to/database.db')
# driver = FalkorDriver(falkor_db=falkordb_client)

# Pass the driver to Graphiti
graphiti = Graphiti(graph_driver=driver)

from graphiti_core import Graphiti
from graphiti_core.driver.kuzu_driver import KuzuDriver

# Create a Kuzu driver
driver = KuzuDriver(db="/tmp/graphiti.kuzu")

# Pass the driver to Graphiti
graphiti = Graphiti(graph_driver=driver)

from graphiti_core import Graphiti
from graphiti_core.driver.neptune_driver import NeptuneDriver

# Create a Neptune driver
driver = NeptuneDriver(
    host='<NEPTUNE_ENDPOINT>',
    aoss_host='<AMAZON_OPENSEARCH_SERVERLESS_HOST>',
    port=8182,      # Optional, defaults to 8182
    aoss_port=443,  # Optional, defaults to 443
)

# Pass the driver to Graphiti
graphiti = Graphiti(graph_driver=driver)

Contributing a new graph backend? See Adding a graph driver.

Graphiti supports Azure OpenAI for both LLM inference and embeddings using Azure's OpenAI v1 API compatibility layer.

from openai import AsyncOpenAI
from graphiti_core import Graphiti
from graphiti_core.llm_client.azure_openai_client import AzureOpenAILLMClient
from graphiti_core.llm_client.config import LLMConfig
from graphiti_core.embedder.azure_openai import AzureOpenAIEmbedderClient

# Initialize Azure OpenAI client using the standard OpenAI client
# with Azure's v1 API endpoint
azure_client = AsyncOpenAI(
    base_url="https://your-resource-name.openai.azure.com/openai/v1/",
    api_key="your-api-key",
)

# Create LLM and Embedder clients
llm_client = AzureOpenAILLMClient(
    azure_client=azure_client,
    config=LLMConfig(model="gpt-5-mini", small_model="gpt-5-mini")  # Your Azure deployment name
)
embedder_client = AzureOpenAIEmbedderClient(
    azure_client=azure_client,
    model="text-embedding-3-small"  # Your Azure embedding deployment name
)

# Initialize Graphiti with Azure OpenAI clients
graphiti = Graphiti(
    "bolt://localhost:7687",
    "neo4j",
    "password",
    llm_client=llm_client,
    embedder=embedder_client,
)

# Now you can use Graphiti with Azure OpenAI

Key Points:

• Use the standard AsyncOpenAI client with Azure's v1 API endpoint format: https://your-resource-name.openai.azure.com/openai/v1/
• The deployment names (e.g., gpt-5-mini, text-embedding-3-small) should match your Azure OpenAI deployment names
• See examples/azure-openai/ for a complete working example

Make sure to replace the placeholder values with your actual Azure OpenAI credentials and deployment names.

Graphiti supports Google's Gemini models for LLM inference, embeddings, and cross-encoding/reranking. To use Gemini, you'll need to configure the LLM client, embedder, and the cross-encoder with your Google API key.

Install Graphiti:

uv add "graphiti-core[google-genai]"

# or

pip install "graphiti-core[google-genai]"

from graphiti_core import Graphiti
from graphiti_core.llm_client.gemini_client import GeminiClient, LLMConfig
from graphiti_core.embedder.gemini import GeminiEmbedder, GeminiEmbedderConfig
from graphiti_core.cross_encoder.gemini_reranker_client import GeminiRerankerClient

# Google API key configuration
api_key = "<your-google-api-key>"

# Initialize Graphiti with Gemini clients
graphiti = Graphiti(
    "bolt://localhost:7687",
    "neo4j",
    "password",
    llm_client=GeminiClient(
        config=LLMConfig(
            api_key=api_key,
            model="gemini-2.0-flash"
        )
    ),
    embedder=GeminiEmbedder(
        config=GeminiEmbedderConfig(
            api_key=api_key,
            embedding_model="embedding-001"
        )
    ),
    cross_encoder=GeminiRerankerClient(
        config=LLMConfig(
            api_key=api_key,
            model="gemini-2.5-flash-lite"
        )
    )
)

# Now you can use Graphiti with Google Gemini for all components

The Gemini reranker uses the gemini-2.5-flash-lite model by default, which is optimized for cost-effective and low-latency classification tasks. It uses the same boolean classification approach as the OpenAI reranker, leveraging Gemini's log probabilities feature to rank passage relevance.

Graphiti can use any OpenAI-compatible /v1 endpoint for LLM inference via OpenAIGenericClient — both hosted providers (DeepSeek, Together, OpenRouter, Fireworks, etc.) and local servers (Ollama, vLLM, llama.cpp, LM Studio). Local servers are ideal for privacy-focused applications or avoiding API costs. The example below uses Ollama; for any other provider, point base_url at its endpoint and set the appropriate api_key and model.

Note: Use OpenAIGenericClient (not OpenAIClient) for these endpoints. It is optimized for local models with a higher default max token limit (16K vs 8K) and handles structured outputs across compatible providers.

Install the models:

ollama pull deepseek-r1:7b # LLM
ollama pull nomic-embed-text # embeddings

from graphiti_core import Graphiti
from graphiti_core.llm_client.config import LLMConfig
from graphiti_core.llm_client.openai_generic_client import OpenAIGenericClient
from graphiti_core.embedder.openai import OpenAIEmbedder, OpenAIEmbedderConfig
from graphiti_core.cross_encoder.openai_reranker_client import OpenAIRerankerClient

# Configure Ollama LLM client
llm_config = LLMConfig(
    api_key="ollama",  # Ollama doesn't require a real API key, but some placeholder is needed
    model="deepseek-r1:7b",
    small_model="deepseek-r1:7b",
    base_url="http://localhost:11434/v1",  # Ollama's OpenAI-compatible endpoint
)

llm_client = OpenAIGenericClient(config=llm_config)

# Initialize Graphiti with Ollama clients
graphiti = Graphiti(
    "bolt://localhost:7687",
    "neo4j",
    "password",
    llm_client=llm_client,
    embedder=OpenAIEmbedder(
        config=OpenAIEmbedderConfig(
            api_key="ollama",  # Placeholder API key
            embedding_model="nomic-embed-text",
            embedding_dim=768,
            base_url="http://localhost:11434/v1",
        )
    ),
    cross_encoder=OpenAIRerankerClient(client=llm_client, config=llm_config),
)

# Now you can use Graphiti with local Ollama models

Ensure Ollama is running (ollama serve) and that you have pulled the models you want to use.

Graphiti depends on structured (JSON) output for entity/edge extraction and deduplication, and works best with models and providers that reliably honor it (OpenAI, Anthropic, Gemini). Reliability varies across OpenAI-compatible providers and especially on smaller or local models, so OpenAIGenericClient exposes a structured_output_mode:

• "json_schema" (default): requests native structured output via response_format. Best on capable models and providers that enforce the schema via constrained decoding.
• "json_object": requests plain-JSON mode and injects the schema into the prompt instead. Use this for providers/models that don't reliably honor json_schema — including some local servers that accept the json_schema request but don't actually constrain output to it, where json_object can be more reliable.

When using smaller or local models:

• Prefer the most capable model you can run. Very small models frequently emit JSON that doesn't match the requested schema, which surfaces as extraction failures.
• Responses wrapped in Markdown ```json code fences are stripped automatically.
• Keep SEMAPHORE_LIMIT low (see above) — local servers and some providers have limited concurrency.

Graphiti lets you override LLM prompt builders per client instance. Override callables must return ChatPrompt (system + user). Response schemas are fixed and not overridable:

from graphiti_core import Graphiti
from graphiti_core.prompts import ChatPrompt, SystemMessage, UserMessage, create_prompt_library

def custom_extract_message(context: dict) -> ChatPrompt:
    return ChatPrompt(
        system=SystemMessage(content='Extract entities for my domain.'),
        user=UserMessage(content=str(context.get('episode_content', ''))),
    )

prompt_library = create_prompt_library({
    'extract_nodes': {
        'extract_message': custom_extract_message,
    },
})

graphiti = Graphiti(..., prompt_library=prompt_library)

Unspecified prompts keep the built-in defaults. Returning list[Message] from an override raises TypeError — migrate with return ChatPrompt(system=..., user=...). Unicode handling is applied via ChatPrompt.as_messages().

The package-level prompt_library keeps the existing list[Message] API. The opt-in prompt customization API uses ChatPromptLibrary and DefaultChatPromptLibrary.

For per-prompt routing across providers, pass an opt-in llm_runtime. Create one LLMTransport for each provider client. Each LLMModel binds a provider model ID to one transport. An unrouted prompt uses the default model. llm_runtime cannot be combined with llm_client or prompt_library. Response schemas stay fixed. You can configure prompt text and model routes:

from typing import Literal

from graphiti_core import Graphiti
from graphiti_core.llm_client.anthropic_client import AnthropicClient
from graphiti_core.llm_client import (
    LLMRuntime,
    LLMTransport,
    OpenAIClient,
    PromptRoutes,
)

OpenAIModels = Literal['gpt-5.1', 'gpt-5-nano']
openai = LLMTransport[OpenAIModels](
    OpenAIClient(),
    models=['gpt-5.1', 'gpt-5-nano'],
)

AnthropicModels = Literal['claude-sonnet-4-5', 'claude-haiku-4-5']
anthropic = LLMTransport[AnthropicModels](
    AnthropicClient(),
    models=['claude-sonnet-4-5', 'claude-haiku-4-5'],
)

main = openai.model('gpt-5.1')
nano = openai.model('gpt-5-nano')
haiku = anthropic.model('claude-haiku-4-5')
runtime = LLMRuntime(
    model=main,
    routes=PromptRoutes(
        extract_nodes=PromptRoutes.ExtractNodes(extract_attributes=nano),
        dedupe_edges=PromptRoutes.DedupeEdges(
            resolve_edge=haiku,
        ),
    ),
)

graphiti = Graphiti(..., llm_runtime=runtime)

LLMModel does not have a small_id field. The runtime passes model_size for legacy compatibility, but this value does not select a smaller model. The selected model ID controls provider calls. A routed prompt uses that model ID. An unrouted prompt uses the default model ID. Route prompts that need a smaller model to that model explicitly. Legacy calls without llm_runtime retain LLMConfig.small_model behavior.

GLiNER2 binds model objects at initialization. It does not support per-prompt model routing. See spec/llm-runtime.md for the runtime API.

JevClient routes node dedupe and edge resolution to TypeSafe Jev. Set JEV_API_KEY before you create the client. The duplicate threshold defaults to 0.7. The contradiction threshold defaults to 0.8.

Jev is a decision model. It answers typed questions with probabilities. It does not generate text. Use a general LLM as the default model. Route only dedupe_nodes.nodes and dedupe_edges.resolve_edge to Jev. Do not use Jev for extraction, summaries, or other generation tasks.

TypeSafe hosts the Jev API. Graphiti sends dedupe prompt content to TypeSafe. See the FalkorDB Jev dedupe example for a runnable example.

from graphiti_core import Graphiti
from graphiti_core.llm_client import (
    JevClient,
    LLMRuntime,
    LLMTransport,
    OpenAIClient,
    PromptRoutes,
    jev_prompt_overrides,
)

openai = LLMTransport(OpenAIClient(), models=['gpt-6-luna'])
jev = LLMTransport(JevClient(), models=['jev-latest'])
jev_dedupe = jev.model('jev-latest', prompt_overrides=jev_prompt_overrides())

runtime = LLMRuntime(
    model=openai.model('gpt-6-luna'),
    routes=PromptRoutes(
        dedupe_nodes=PromptRoutes.DedupeNodes(nodes=jev_dedupe),
        dedupe_edges=PromptRoutes.DedupeEdges(resolve_edge=jev_dedupe),
    ),
)
graphiti = Graphiti(..., llm_runtime=runtime)

OpenAIDecisionsClient routes node dedupe and edge resolution to the OpenAI Decisions API. Use a general LLM as the default model. A decision model answers typed questions with probabilities and does not generate text. Do not use it for extraction, summaries, or other generation tasks.

See the OpenAI Decisions FalkorDB example for a runnable example.

• Guides and API documentation.
• Quick Start
• Building an agent with LangChain's LangGraph and Graphiti

Graphiti collects anonymous usage statistics to help us understand how the framework is being used and improve it for everyone. We believe transparency is important, so here's exactly what we collect and why.

When you initialize a Graphiti instance, we collect:

• Anonymous identifier: A randomly generated UUID stored locally in ~/.cache/graphiti/telemetry_anon_id
• System information: Operating system, Python version, and system architecture
• Graphiti version: The version you're using
• Configuration choices:
  • LLM provider type (OpenAI, Azure, Anthropic, etc.)
  • Database backend (Neo4j, FalkorDB, Kuzu, Amazon Neptune Database or Neptune Analytics)
  • Embedder provider (OpenAI, Azure, Voyage, etc.)

We are committed to protecting your privacy. We never collect:

• Personal information or identifiers
• API keys or credentials
• Your actual data, queries, or graph content
• IP addresses or hostnames
• File paths or system-specific information
• Any content from your episodes, nodes, or edges

This information helps us:

• Understand which configurations are most popular to prioritize support and testing
• Identify which LLM and database providers to focus development efforts on
• Track adoption patterns to guide our roadmap
• Ensure compatibility across different Python versions and operating systems

By sharing this anonymous information, you help us make Graphiti better for everyone in the community.

The Telemetry code may be found here.

Telemetry is opt-out and can be disabled at any time. To disable telemetry collection:

Option 1: Environment Variable

export GRAPHITI_TELEMETRY_ENABLED=false

Option 2: Set in your shell profile

# For bash users (~/.bashrc or ~/.bash_profile)
echo 'export GRAPHITI_TELEMETRY_ENABLED=false' >> ~/.bashrc

# For zsh users (~/.zshrc)
echo 'export GRAPHITI_TELEMETRY_ENABLED=false' >> ~/.zshrc

Option 3: Set for a specific Python session

import os

os.environ['GRAPHITI_TELEMETRY_ENABLED'] = 'false'

# Then initialize Graphiti as usual
from graphiti_core import Graphiti

graphiti = Graphiti(...)

Telemetry is automatically disabled during test runs (when pytest is detected).

• Telemetry uses PostHog for anonymous analytics collection
• All telemetry operations are designed to fail silently - they will never interrupt your application or affect Graphiti functionality
• The anonymous ID is stored locally and is not tied to any personal information

We encourage and appreciate all forms of contributions, whether it's code, documentation, addressing GitHub Issues, or helping others on GitHub. For detailed guidelines on code contributions, please refer to CONTRIBUTING.

Open a GitHub issue to ask questions, report bugs, or discuss Graphiti.