Letta (formerly MemGPT) is an open framework for building stateful AI agents with persistent memory. The original MemGPT paper proposed treating an LLM's context window like main memory in an operating system: the agent can page memories in and out, archive long-term context, and recall it later. Letta packages this into a library and an open server with a managed cloud option. Apache 2.0. Letta matters because the deepest agent lock-in is accumulated memory. Closed agents (Claude Code, ChatGPT, Cursor) build memories about you that are typically not exportable; switching to a different agent means starting over. Letta is the leading open answer: an agent backed by a memory store you control, with the memory itself in your database. Compared to siblings: mem0 (similar mission, different memory-tier design), Zep (also open, focused on developer agents), the memory features baked into closed agents (the same idea but server-side and locked to the vendor). Production-readiness: beta. The framework works and ships; production deployments at scale are emerging. Used by developers building stateful chatbots, personal-AI assistants, and research agents. Strategic significance: if open agents are going to displace closed ones, the memory layer is one of the hardest gaps to close, and Letta is the best current open bet on closing it.
The Stack · Retrieval and Memory · Open source
Letta (formerly MemGPT)
Memory-stateful agent framework; OS-like memory hierarchy; persistent context across sessions.
Sources
- Letta https://letta.com/
- Letta on GitHub https://github.com/letta-ai/letta
- MemGPT Paper (Towards LLMs as Operating Systems) https://arxiv.org/abs/2310.08560
Want a follow-up? Ask the chat about Letta (formerly MemGPT) in context. It will compare to siblings at the same layer and ground every claim in the wiki.
Other projects at the Retrieval and Memory layer
8 siblings · ordered open first
- Qdrant Open source
Rust-native open-source vector database; production-grade similarity search and filtering.
- Weaviate Open source
Go-based open-source vector database; built-in modules for embedding and reranking.
- Chroma Open source
Python-first open-source vector database; popular default for AI-app development.
- Milvus Open source
Go-based open-source vector database; large-scale production deployments; Zilliz commercial offering.
- LanceDB Open source
Rust-native open vector database; serverless; columnar Lance format underneath.
- pgvector Open source
Vector similarity search inside PostgreSQL; lets existing Postgres deployments add RAG without a separate database.
- mem0 Open source
Memory layer for AI agents; vector-store-agnostic; popular for adding persistent user state to chat apps.
- Pinecone Proprietary
Closed managed vector database; fast and operationally simple; no self-host option at most tiers.