Graphiti — Temporal AI Knowledge Graph by Zep

What it is

Graphiti is a Python library by Zep that builds dynamic knowledge graphs from AI agent conversations and unstructured data. Unlike static knowledge graphs, Graphiti tracks entity changes over time, resolves contradictions when new information conflicts with old, and provides temporal-aware queries. It stores entities and relationships in Neo4j with timestamps for every fact.

Graphiti is designed for AI engineers building agents that need persistent, evolving memory with the ability to query what was true at any point in time.

How it saves time or tokens

Traditional AI memory (vector stores, chat history) loses temporal context. If a user says 'I work at Google' in January and 'I just started at Meta' in March, a vector store returns both facts without knowing which is current. Graphiti tracks these changes: it marks the Google fact as superseded and the Meta fact as current. This eliminates contradictions in agent memory and reduces the tokens spent resolving conflicting context.

How to use

1. Install Graphiti:

pip install graphiti-core

1. Initialize with Neo4j and add episodes:

from graphiti_core import Graphiti
from graphiti_core.llm_client import AnthropicClient

graphiti = Graphiti(
    uri='bolt://localhost:7687',
    llm_client=AnthropicClient(),
)

await graphiti.build_indices()

# Add conversation episodes
await graphiti.add_episode(
    name='user_chat_1',
    episode_body='User said: I work at Google as a senior engineer.',
    source_description='chat',
)

await graphiti.add_episode(
    name='user_chat_2',
    episode_body='User said: I just started at Meta last week.',
    source_description='chat',
)

1. Query the graph:

results = await graphiti.search('Where does the user work?')
# Returns: Meta (current), with Google marked as previous

Example

Temporal querying to understand entity changes:

# Query current state
current = await graphiti.search('user employment')
print(current)  # User works at Meta (since March 2026)

# Query historical state
from datetime import datetime
historical = await graphiti.search(
    'user employment',
    as_of=datetime(2026, 2, 1),
)
print(historical)  # User works at Google (as of Feb 2026)

# The graph automatically:
# - Created a 'User' entity node
# - Created 'Google' and 'Meta' organization nodes
# - Created 'WORKS_AT' relationships with timestamps
# - Marked the Google relationship as superseded

Related on TokRepo

• AI memory tools — Browse persistent memory solutions for AI agents
• Knowledge graph tools — Explore knowledge graph tools

Common pitfalls

• Not running Neo4j before starting Graphiti. Graphiti requires a Neo4j instance for graph storage. Start Neo4j with Docker: docker run -p 7687:7687 -p 7474:7474 neo4j.
• Adding too many episodes at once without letting the graph consolidate. Graphiti resolves contradictions during episode ingestion. Large batch ingestion may produce unexpected entity merges.
• Expecting Graphiti to work without an LLM. Graphiti uses an LLM (Claude or GPT) to extract entities and relationships from unstructured text. An LLM API key is required.