Letta

Letta is a code-first AI agent harness and SDK for developers building stateful agents with git-versioned memory, run on Letta's cloud or self-hosted.

Letta gives an AI agent a memory that lasts across sessions, so it keeps learning about a user, a codebase or a team instead of starting every conversation blank. Letta Inc, an independent San Francisco lab that grew out of the MemGPT research project, builds it. Letta Agent is the self-improving agent at its centre. One harness runs behind the app, the CLI, the Agent SDK and the Letta API. Memory lives in MemFS, a set of Markdown files committed to git on every edit, so prompts and skills carry a full version history. The Letta Harness, formerly Letta Code, is the coding agent in the terminal and in the desktop and web app. Subagents take parts of a task and report back to the main agent, and permission modes ask before shell commands or file edits. Letta Auto picks a frontier model for each job, and any agent can switch models mid-conversation. Agents reach tools over MCP, take messages from Slack, Discord and Telegram, and answer GitHub issue and pull request comments through a GitHub Action. Letta suits developers who want an agent defined in code and kept for months, not rebuilt one prompt at a time. Individual coders start in the app or CLI on a personal plan, using their own API keys or Letta Auto usage. Product teams that embed agents in their own software build on the Agent SDK and the Letta API, which bill by usage: active agents, tool execution time and model tokens. Teams share agents under per-seat plans with access control, and larger companies buy by quote with single sign-on and role-based access. Organizations with strict data rules run the Letta App Server on their own machines, with no Letta account required. A coding session shows the pattern: one agent fixes a bug, then reviews recent pull requests in parallel through its subagents. The same agent can join a team's Slack workspace, take requests there and run tasks on a schedule. Developers building on the API give each customer a stateful agent inside a web or React Native app, with memory that survives between visits. Letta fits teams already working in GitHub and chat that want an agent which improves with use rather than resetting at every session.