How to Migrate from Mem0 to REM Labs
Mem0 is a solid developer tool. But if you need higher recall accuracy, multi-signal retrieval beyond vector search, or features like Memory Synthesis -- it is time to switch. This guide covers the one-command migration, what gets transferred, and why the benchmark numbers matter.
The Benchmark Case
LongMemEval (ICLR 2025) is the standard benchmark for long-term AI memory systems. It tests 500 questions across six categories: single-session recall, preference extraction, knowledge updates, temporal reasoning, and multi-session synthesis. Here is how the two systems compare:
| System | LongMemEval Score | Questions Correct |
|---|---|---|
| REM Labs | 90% | 450 / 500 |
| Mem0 | 66.9% | ~335 / 500 |
| ChatGPT Memory | 52.9% | ~265 / 500 |
The 30-point gap is not about fine-tuning or parameter optimization. It is an architectural difference. Mem0 uses vector-only retrieval -- embeddings and cosine similarity. REM Labs runs four retrieval paths in parallel (vector, full-text, entity graph, structured metadata), fuses the results, and applies neural reranking. The full technical breakdown is in our LongMemEval deep dive.
Step 1: Get a REM Labs API Key
Sign up at remlabs.ai/console or run:
This gives you an API key on the free tier. You can migrate and test before committing to a paid plan.
Step 2: Run the Migration
One command exports all memories from your Mem0 account and imports them into REM Labs:
The migration tool handles both v1 and v2 Mem0 schemas automatically. For all available options, run rem migrate --help.
What Gets Migrated
- Memory content -- every stored memory value is transferred
- Metadata -- tags, timestamps, and custom metadata fields are preserved
- Namespaces -- Mem0 user/agent IDs are mapped to REM Labs namespaces
- Timestamps -- original creation and update times are preserved for temporal ordering
What Gets Added
After migration, your memories are re-indexed with REM's full retrieval stack. Each memory gains:
- Full-text index entry -- enabling exact keyword, proper noun, and acronym search that vector-only systems miss
- Entity extraction -- people, places, organizations, and concepts are extracted and added to the knowledge graph
- Multi-signal search -- queries now run four parallel retrieval paths instead of one
Step 3: Update Your Code
Python
Node.js
Step 4: Verify
After migration, confirm your memories transferred correctly:
Non-Destructive by Default
The migration is non-destructive. Your data in Mem0 is never modified or deleted. You can run both systems in parallel during a transition period, routing reads to REM Labs while keeping Mem0 as a fallback.
Migration source code: The migration tool is open source. Review the implementation at github.com/remlabs/agent-apis.
Why Switch
Beyond the benchmark numbers, here is what you gain:
- Full-text search -- the single largest accuracy improvement. Proper nouns, exact phrases, and technical terms that embeddings miss are retrieved with high precision.
- Knowledge graph -- entity extraction and graph traversal for structured relationship queries.
- Memory Synthesis -- a deep-synthesis layer with 9 consolidation strategies that connects signals across your memory corpus and surfaces insights. Run it via
POST /v1/dream/run. - 40+ integrations -- LangChain, CrewAI, Obsidian, MCP for Claude Code and Cursor, Slack, Notion, and more.
- Self-hostable -- Docker images and Kubernetes Helm charts available for enterprise deployments.
Migrate in one command
Non-destructive. Full metadata preserved. 90% recall accuracy.
Get started free →