What Is Letta (formerly MemGPT)?

Letta, formerly known as MemGPT, is an open-source framework for building stateful AI agents with long-term memory. Its central idea is that agents should manage their own memory and persist context across interactions, rather than starting fresh each time. This article explains what Letta is, where it came from, and how its memory model works.

What Letta Is

Letta is an open-source, model-agnostic framework for building, deploying, and managing stateful AI agents. The word "stateful" is key: a Letta agent retains memory across conversations and sessions, so it can accumulate knowledge over time and reason over what it has learned. This sets it apart from stateless setups where each request is handled in isolation.

The project grew out of MemGPT, a research project from UC Berkeley's AI research community. MemGPT introduced the concept of virtual context management—an operating-system-inspired approach to giving language models effectively unlimited context. Today, MemGPT refers to that original design pattern, while Letta is the name of the framework and the company building around it.

The Memory Architecture

Letta's most distinctive feature is its memory model, which is inspired by how an operating system manages memory. The agent works with a main context, which is the working memory the model sees on every turn, much like RAM. Beyond that, recent message history lives in recall storage that can be searched on demand, and long-term knowledge is kept in archival storage that the agent can query when needed.

What makes this approach notable is that the agent manages its own memory through tool use, rather than relying solely on an external retrieval pipeline bolted on by the developer. The agent can decide what to move into long-term storage, what to bring back into working memory, and how to edit its own memory over time. This self-editing memory is the core innovation carried over from MemGPT.

What You Can Build

With persistent memory and reasoning, Letta agents are suited to applications where continuity matters. They can maintain long-term memory across many interactions, reason over accumulated knowledge, integrate with external tools and APIs, and operate either autonomously or alongside other agents. This makes them a fit for use cases like ongoing assistants, research helpers, and task management where remembering past context improves results.

Because Letta is open source and model-agnostic, developers can run it with different underlying models and self-host it. The framework provides the structure for defining agents, their memory, and their tools, and it exposes ways to inspect and manage agent state, which supports transparency into what an agent remembers and why.

Who It Is For

Letta appeals to developers building agents that need genuine long-term memory and continuity rather than one-off responses. Its operating-system-inspired memory model is a strong fit when an agent must learn and improve over many sessions. As an open-source project, it can be adopted freely and self-hosted, with the option of using managed tooling from the project's maintainers.