Letta Code Review 2026

Memory-first coding agent that persists across sessions and is portable across models — from the team behind MemGPT.

Letta Code is a memory-first coding agent built by the Letta team (creators of MemGPT and pioneers of LLM memory research). The pitch is direct: most coding agents work in independent sessions and lose all context when the conversation resets. Letta Code instead works with a persisted agent that remembers everything you've done, learns continuously from corrections and feedback, and stays portable across models — switch from Claude to GPT to an open-weight model without losing the agent's memory or skills. The harness includes MemFS (a virtual filesystem memory layer), Skills (reusable agent capabilities), Subagents for parallel work, and pluggable Providers so you can connect any model or coding plan. Headless mode and a CLI ship for terminal use, a desktop app for richer interactions, and remote/mobile access so you can check in on long-running agent work from your phone. A GitHub Action lets the same agent operate inside CI.

Key capabilities at a glance: Persisted agent memory across sessions; Model-portable: switch providers without losing memory; MemFS virtual filesystem memory layer; Skills, Subagents, and Hooks for extensibility; Desktop app, CLI, headless mode, GitHub Action; Letta Code SDK for building custom agents.