Letta
An AI research lab building stateful agents with memory-native learning. Agents remember everything, adapt continuously, and evolve their capabilities through experience.
Letta is an open-source alternative to:
Letta is an AI research lab in San Francisco focused on building machines that learn. Their vision is a future where humans coexist with digital people: experiential agents that remember everything, learn continuously, and improve themselves over time.
Key Features
- Stateful AI Agents: Create agents with persistent memory that evolves through experience
- Memory-Native Learning: Agents learn through specialized memory models trained with memory-native reinforcement learning
- Git-Based Memory: Programmatic context management with git-based versioning for agent memory
- Offline Learning: Sleep-time compute allows agents to reason and learn during idle periods
- Self-Improving Capabilities: Agents continuously refine their identity, memory, and capabilities based on interactions
- Research-Backed: Built on foundational work like MemGPT (LLM Operating System) from UC Berkeley's Sky Computing Lab
Tech / Self-Hosting
Letta's research ships as software through the Letta Agent, available as a npm package (@letta-ai/letta-code). The technology is designed for enterprise and research applications, with case studies from companies like Bilt, 11x, Kognitos, and Hunt Club. The team consists of researchers and engineers from UC Berkeley's Sky Computing Lab, advised by industry leaders including Ion Stoica and Joey Gonzalez.