Letta Code is a memory-first CLI coding agent built on the Letta platform. Unlike traditional coding agents that reset with every session, Letta Code uses long-lived stateful agents that persist across sessions, remembering your codebase, preferences, and past decisions. The interface is intentionally familiar to Claude Code and Codex users — run it from the terminal to code, edit files, run programs, and manage projects.
The core innovation is persistent memory through structured memory blocks (persona, human, project, skills) stored in git-backed context repositories. This enables versioning, branching, and diffing of agent knowledge. Sleep-time reflection subagents periodically consolidate and improve memory in the background without interrupting your workflow.
Fully model-agnostic — the #1 model-agnostic open-source harness on Terminal-Bench, performing comparably to provider-specific tools when using the same models. Supports Claude Sonnet/Opus, GPT/Codex, Gemini, GLM, Kimi, and more. Switch models with /model while keeping all memory intact.
Skill learning lets agents capture workflow patterns from experience for reuse. Subagent orchestration delegates specialized tasks. The /init command bootstraps memory from your project and even imports context from prior Claude Code or Codex sessions.
Pricing: Free tier, Pro $20/mo (20 agents), Max Lite $100/mo (50 agents, frontier models), Max $200/mo (power users, 20x limits), API $20/mo + $0.10/active agent for teams. Install via npm. The underlying Letta platform is open-source (22K+ stars) with Python and TypeScript SDKs for building custom stateful agent applications.