Letta

Letta is an open-source AI agent platform built by the creators of MemGPT at UC Berkeley's Sky Computing Lab. It provides both a CLI coding agent (Letta Code) and a full-featured API for building stateful agents that persist memory across sessions, learn from experience, and remain portable across model providers. The project is licensed under Apache 2.0 and actively maintained on GitHub.

What It Is

Letta sits in the category of stateful agent frameworks — tools that go beyond single-session LLM calls to give agents persistent, evolving memory. Where most LLM applications reset context on every request, Letta agents carry memory blocks that update over time, enabling them to remember users, accumulate skills, and improve from prior interactions. The platform exposes this capability through two main surfaces: the Letta Code CLI agent (installable via npm) and the Letta API with Python and TypeScript SDKs.

Architecture: Memory-First Design

The core architectural idea in Letta is that memory is a first-class citizen, not an afterthought. Key design elements include:

• Memory blocks: Structured context segments (e.g., "human" and "persona" blocks) that agents read and write during inference
• Dream agents: Background memory agents that transform prompts, context, and skills over time during idle periods — what the team calls "sleep-time compute"
• Context Repositories: Git-based versioning for agent memory, enabling programmatic context management and rollback
• Skill learning: Agents dynamically acquire and store reusable skills through experience
• Model portability: Agent memories, conversations, and experiences can be transferred across any model provider

Letta Code: The CLI Agent

Letta Code is the flagship end-user product — a memory-first coding harness that runs locally in the terminal. It installs via npm install -g @letta-ai/letta-code (requires Node.js 18+) and launches an agent that can access the local filesystem, execute tasks on the user's computer, and accumulate skills over time. Desktop app downloads are available for macOS (Apple Silicon and Intel), Windows, and Linux. The GitHub README notes that Letta Code supports skills and subagents, and bundles pre-built capabilities for advanced memory and continual learning.

Research Lineage and Team

Letta was founded by Charles Packer (PhD, UC Berkeley) and Sarah Wooders (PhD, UC Berkeley), the creators of MemGPT. The company describes itself as an AI lab building "machines that learn," with advisors including Ion Stoica and Joseph Gonzalez (both UC Berkeley professors and co-creators of Apache Spark and Ray). The About page lists backing from Jeff Dean (Chief Scientist at Google DeepMind), Clem Delangue (CEO of HuggingFace), and other AI infrastructure leaders.

Update: v0.16.8 and Active Development

The GitHub repository (letta-ai/letta) shows the latest release as v0.16.8, published May 14, 2026, with the repository last updated May 20, 2026. The project has accumulated over 22,800 stars and 2,400 forks. Recent blog posts from the Letta team cover Context Repositories (git-based memory), Skill Learning for CLI agents, sleep-time compute, and Context-Bench — a benchmark for evaluating LLMs on agentic context engineering tasks. This activity signals an actively evolving research-to-product pipeline.

Open-Source Deployment Model

The core Letta platform is open source under Apache License 2.0, meaning it can be self-hosted, modified, and redistributed freely. Developers can run agents entirely locally using the CLI, or connect to the hosted Letta API for cloud-based stateful agent infrastructure. The API offers Python and TypeScript SDKs with a straightforward agent creation and messaging interface. The project welcomes contributions and maintains a Discord community and developer forum.