Migrating from Mem0 to Cognee

Migration steps

1. 1

   Export your memories from Mem0

   Use the Mem0 SDK or platform export to pull every memory you want to keep (per user, agent, or all) and save the result as JSON. The Cognee adapter accepts a plain list of memory records or a dict with a results, memories, or items key, which matches what Mem0's list and export endpoints return.

   import json
   from mem0 import MemoryClient   # or Memory() for the OSS package
   
   client = MemoryClient()
   records = client.get_all(user_id="alice")   # repeat per user/agent as needed
   
   with open("mem0_export.json", "w") as f:
       json.dump(records, f)

2. 2

   Install Cognee and pick your storage

   pip install cognee gives you embedded defaults that need no infrastructure. Set LLM_API_KEY, and optionally point the graph and vector layers at PostgreSQL, Neo4j, or Neptune through environment variables when you are ready for production.

   pip install cognee
   export LLM_API_KEY="..."   # OpenAI, Anthropic, or a local provider

3. 3

   Import with Mem0Source

   Pass the export to cognee.remember through the Mem0Source adapter. The default mode re-derives the graph by running Cognee's own extraction over the raw memories; use preserve to map Mem0's facts straight into the graph with zero LLM calls, or hybrid for both.

   import asyncio
   import cognee
   from cognee.migration import Mem0Source
   
   async def main():
       await cognee.remember(
           Mem0Source("mem0_export.json", mode="re-derive"),
           dataset_name="customer_memory",
       )
   
   asyncio.run(main())

4. 4

   Verify the import

   Ask the questions your agents will ask. GRAPH_COMPLETION answers from the graph; CHUNKS returns the raw records so you can spot-check that nothing was dropped.

   from cognee import SearchType
   
   results = await cognee.search(
       "What does Alice prefer?",
       query_type=SearchType.GRAPH_COMPLETION,
       datasets=["customer_memory"],
   )

5. 5

   Point your agents at Cognee

   Replace Mem0 add/search calls with cognee.remember and cognee.recall in the Python SDK, use the TypeScript client (@cognee/cognee-ts), or connect Claude Code, Cursor, and other MCP clients to the Cognee MCP server. Keep Mem0 running in parallel until the new answers match.

Mode choice is the main decision. re-derive (default) throws away Mem0's extracted facts and rebuilds the graph from the raw memory text with Cognee's pipeline, which costs LLM tokens and produces the richest graph. preserve keeps Mem0's facts as-is with no LLM calls, which is fast and cheap but flat. hybrid does both. Start with preserve on a sample to check the shape, then re-derive the full set.

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.