Graphiti — Real-Time Knowledge Graphs for AI Agents

What it is

Graphiti is an open-source library by Zep that builds real-time knowledge graphs for AI agents. Unlike static knowledge bases, Graphiti continuously ingests new information, extracts entities and relationships, and maintains temporal awareness -- it knows when facts were added, updated, or became stale. The knowledge graph serves as long-term structured memory for agents that need to remember and reason over accumulated information.

Developers building AI agents that need persistent, structured memory beyond simple conversation history benefit most. Agents handling customer support, research, or multi-session workflows need to remember entities, relationships, and how they change over time.

How it saves time or tokens

Without Graphiti, agents typically store memory as flat text, which requires embedding search and often returns irrelevant context. Graphiti's graph structure enables precise entity lookups and relationship traversals. An agent can ask 'what is the status of Project X?' and get the current answer from the graph rather than searching through thousands of conversation turns. The temporal layer means outdated information does not pollute queries -- only the latest state is returned unless history is specifically requested.

How to use

1. Install Graphiti:

pip install graphiti-core

1. Initialize the graph and add episodes:

from graphiti_core import Graphiti

graph = Graphiti(neo4j_uri='bolt://localhost:7687', neo4j_user='neo4j', neo4j_password='password')

await graph.add_episode(
    name='meeting_notes',
    episode_body='Alice joined the team as lead engineer on April 10. The project deadline is May 15.',
    source='meeting'
)

1. Query the graph for entities and relationships:

results = await graph.search('What role does Alice have?')
print(results)  # Returns: Alice -> lead engineer, added April 10

Example

# Add multiple episodes over time
await graph.add_episode(
    name='update_1',
    episode_body='Project deadline moved from May 15 to June 1 due to scope changes.',
    source='slack'
)

# Graphiti updates the temporal state automatically
results = await graph.search('project deadline')
# Returns: June 1 (latest), with history showing May 15 was previous

# Entity-specific queries
entities = await graph.get_entity('Alice')
# Returns all relationships: role, team, projects, timeline

Related on TokRepo

• AI Memory Providers -- Graphiti deep-dive on TokRepo
• Knowledge Graph Tools -- Knowledge graph tools for AI agents

Common pitfalls

• Graphiti requires Neo4j as the graph database backend. Set up Neo4j before using Graphiti -- the Docker image is the quickest path.
• Entity extraction quality depends on the LLM used for processing episodes. Use a capable model (Claude, GPT-4) for accurate entity and relationship extraction.
• Large volumes of episodes can slow down graph updates. Batch episode ingestion during off-peak times for production workloads.