Migrating from Graphiti to Cognee

Migration steps

1. 1

   Dump your Graphiti graph to JSON

   Query your Graphiti store for entity nodes, entity edges (facts), and episodic nodes, and write them out as a dict with nodes, edges, and episodes keys. The adapter reads the same shape as Zep's graph export, so a straightforward Cypher dump is enough.

   import json
   from neo4j import GraphDatabase
   
   driver = GraphDatabase.driver("bolt://localhost:7687", auth=("neo4j", "..."))
   
   def rows(query):
       with driver.session() as s:
           return [r.data() for r in s.run(query)]
   
   export = {
       "nodes": rows("MATCH (n:Entity) RETURN n.uuid AS uuid, n.name AS name, n.summary AS summary, labels(n) AS labels"),
       "edges": rows("MATCH (a:Entity)-[r:RELATES_TO]->(b:Entity) RETURN r.uuid AS uuid, a.uuid AS source_node_uuid, b.uuid AS target_node_uuid, r.name AS name, r.fact AS fact, r.valid_at AS valid_at, r.invalid_at AS invalid_at"),
       "episodes": rows("MATCH (e:Episodic) RETURN e.uuid AS uuid, e.content AS content, e.source AS source, e.created_at AS created_at"),
   }
   
   with open("graphiti_dump.json", "w") as f:
       json.dump(export, f, default=str)

2. 2

   Install Cognee

   pip install cognee and set LLM_API_KEY. The embedded defaults (SQLite, LanceDB, Ladybug) are enough for the import; point Cognee at your existing Neo4j with the neo4j extra if you want to keep using it.

   pip install cognee            # or: pip install "cognee[neo4j]"
   export LLM_API_KEY="..."

3. 3

   Import with GraphitiSource

   preserve maps Graphiti's entities and facts directly into Cognee's graph with no LLM calls. hybrid also runs Cognee's extraction over the episodes, adding summaries, chunks, and a vector index so every retrieval mode works.

   import asyncio
   import cognee
   from cognee.migration import GraphitiSource
   
   async def main():
       await cognee.remember(
           GraphitiSource("graphiti_dump.json", mode="hybrid"),
           dataset_name="agent_memory",
       )
   
   asyncio.run(main())

4. 4

   Verify

   Run the temporal and relational questions your agents depend on and compare them with Graphiti's answers. CYPHER search lets you inspect the imported graph directly if Cognee is backed by Neo4j.

   from cognee import SearchType
   
   await cognee.search(
       "Which vendor did the team switch to, and when?",
       query_type=SearchType.TEMPORAL,
       datasets=["agent_memory"],
   )

5. 5

   Reconnect your agents

   Replace Graphiti add_episode and search calls with cognee.remember and cognee.recall, or connect through the Cognee MCP server or the TypeScript client. Keep Graphiti running in parallel until the answers match.

Graphiti has already done entity and fact extraction, so preserve mode is a lossless, zero-LLM copy and the right first run. hybrid (recommended for the final import) keeps that graph and adds Cognee's chunks, summaries, and vector index over the episodes. re-derive ignores Graphiti's graph entirely and rebuilds from the episode text.

The import adapters live in cognee.migration and are documented at docs.cognee.ai. If your export does not match the shape the adapter expects, open an issue on GitHub with a redacted sample.