Cognee vs Letta

Cognee

Builds an LLM-derived knowledge graph alongside embeddings, so recall follows relationships, not just vector similarity.

• Self-hostable Python SDK
• Recall follows concept relationships
• Bring your own LLM/embedding provider

• Newer, smaller ecosystem
• Cognify pipeline adds LLM cost
• Self-host setup overhead

Letta

The productized MemGPT successor — agents edit their own memory blocks (memory-as-OS) to manage a finite context window.

• Self-editing agent memory
• Open source, model-agnostic
• REST APIs + multi-language SDKs
• Reference architecture for memory

• Developer-focused, not no-code
• Memory model has a learning curve
• Younger, evolving framework

Pick Cognee if you need Knowledge graph, RAG pipeline, Embeddings, and Vector search.

• Knowledge graph (secondary capability)
• RAG pipeline (secondary capability)
• Embeddings (secondary capability)
• Vector search (secondary capability)

Pick Letta if you need Agent framework, Tool / function calling, Multi-agent orchestration, and App / agent deployment.

• Agent framework (primary capability)
• Tool / function calling (secondary capability)
• Multi-agent orchestration (secondary capability)
• App / agent deployment (secondary capability)