MemGPT is an 🏷️#open-source project for giving 📝large language model agents self-editing memory, begun as a research project at UC Berkeley. The website of 📝Letta lists "MemGPT: The LLM Operating System" under October 2023 as "the original work on virtual context management — where Letta began." In a September 23, 2024 post titled "MemGPT Is Now Part of Letta," the team wrote that "MemGPT should refer to the original agent design pattern described in the research paper (empowering LLMs with self-editing memory tools), and use the name Letta to refer to the agent framework." The repository at github.com/letta-ai/letta carries the Apache License 2.0, and its README calls the project "Letta (f.k.a. MemGPT)."
According to that post, MemGPT was inspired by an effort to improve the memory systems of Discord chatbots, and the repository grew from the paper to support agents more generally. The post says the name had come to mean three things: the "LLM OS" techniques in the paper, an archetype of LLM-driven chatbots with self-editing memory, and the open-source agent framework. Letta, which its website says was founded by the creators of MemGPT from UC Berkeley's Sky Computing Lab, said it would continue to maintain the open-source repository, moving the Python package to letta and the Docker image to letta/letta-server. The repository's README now points active development to the letta-ai/letta-code repository and keeps the retired Letta V1 API server on an archive branch.
Key Features
- Self-editing memory tools — The MemGPT design pattern gives an LLM tools to edit its own memory, according to Letta.
- Virtual context management — Letta's website calls MemGPT the original work on virtual context management.
- Custom tools and data sources — The repository grew to support custom tools, data sources and memory classes for agents.
- Stateful agent services — The framework added the ability to deploy agents and persist them as stateful services.