Letta Code

Letta Code is a stateful agent harness for creating agents that carry memory, identity, and a sense of experience over time. Agents learn and evolve over long horizons by rewriting their own memory, skills, prompts, and even the harness itself through mods. It can run interactively or power always-on agents that work proactively, and you reach the same agent from a local CLI, a desktop app for macOS, Windows, and Linux, a browser including mobile, and messaging integrations such as Telegram, Slack, and Discord.

Agents are designed to be self-configuring: you can ask an agent to set up its own skills, hooks, and permissions. All context, including memory blocks, is tracked with Git in a MemFS layer that can sync to a GitHub repository. Agents rewrite their context to adapt, with periodic dreaming on a schedule, message search across agents and conversations, and skills loaded at global, project, and agent scope. Built-in subagents run in the background, and any agent can call another agent, including itself, as a subagent.

Agent state, including memory, identity, and conversations, is stored in Letta Cloud by default, while Letta Code itself runs on a laptop, a cloud sandbox, GitHub Actions, or a remote computer, so one agent can move across multiple machines. Running locally is also an option. You connect your own model keys and swap models during a session. External skills install from a GitHub URL, ClawHub, or the Hermes Skills Hub.