MemGPT is an AI agent research direction and software project advanced by Letta. The official website positions it as "teaching machines to learn how to learn": creating stateful agents that can remember everything, continuously learn, and improve themselves over time.
Purpose and positioning: According to the official website introduction, Letta is an AI research lab located in San Francisco, with the vision of humans coexisting with "digital people"—experiential agents that remember everything, continuously learn, and self-improve. The official website emphasizes that "learning in token space" is key to building AI agents that truly improve over time, and proposes a series of principles around how agents manage context to learn from experience.
Main directions and capabilities: The research content listed on the official website includes: memory models trained with memory-native RL to drive agent learning; a set of principles for managing AI agent context to learn from experience; rearchitecting the way agent memory works, using programmatic context management and git-based versioning; having agents reason over context during idle time rather than at inference time (sleep-time compute); and the initial virtual context management work—the starting point of MemGPT. The official website also mentions that the research is delivered as software: Letta Agent is a self-improving AI agent whose memory, identity, and capabilities evolve with experience.
How to use: The installation command provided on the official website is $ npm i -g @letta-ai/letta-code. Apart from this command, the official website does not provide more detailed usage steps or configuration instructions, and none are currently provided.
Applicable scenarios: The case studies showcased on the official website include Bilt, 11x, Kognitos, and Hunt Club. In addition, the official website lists coverage of the project by multiple media outlets, such as WIRED, Fast Company, TechCrunch, etc., with topics involving large language models and AI agents, sleep-time compute, and more. These cases and reports can serve as references for understanding its application directions.
Background and team: The official website states that the project was founded by the creators of MemGPT at the UC Berkeley Sky Computing Lab (the birthplace of Spark and Ray), with advisors Ion Stoica and Joey Gonzalez, and investmentfrom Jeff Dean, Clem Delangue, Cris Valenzuela, Jordan Tigani, Tristan Handy, Barry McCardel, Robert Nishihara, Sunflower Capital, Essence VC, and Felicis Ventures. The team is composed of researchers and engineers.
Limitations and missing information: The official website does not state the specific pricing model, prices, platform support scope, or detailed feature list, and none are currently provided; the above content is only the facts covered by the official website introduction, and does not expand on further unverified capabilities or effects.
