Letta Code is a memory-first coding agent that enhances how coding agents interact and learn over time. Unlike traditional agents that operate in isolated sessions, Letta Code maintains a persistent state across sessions, allowing it to accumulate knowledge and improve based on user interactions. It ranks as the top model-agnostic open-source harness on TerminalBench, delivering performance on par with proprietary solutions like Claude Code, Gemini CLI, and Codex CLI. A key feature is its ability to learn from experience through a process called agentic context engineering. When users initiate a session with the `/init` command, the agent analyzes the local codebase, creating memories that help it adapt to specific user projects. Users can further guide the learning process using the `/remember` command. Skill learning is another vital aspect, enabling agents to retain and apply knowledge from complex tasks, which can be referenced for similar future tasks. These skills, stored as markdown files, can be versioned in git repositories and shared with other agents. Letta Code also excels in memory management. Agents can access past conversations, including those with other agents, through the Letta API. The `/search` command simplifies finding relevant messages, supporting various search methods like vector and full-text. This level of statefulness and learning capability gives Letta Code a competitive edge over other harnesses, making it a robust choice for developers looking for efficient and intelligent coding assistance.