Letta Code launches as memory-first coding agent from MemGPT team

Letta AI — the company that emerged from the MemGPT research project at UC Berkeley — released Letta Code in December 2025 as an application of its persistent memory architecture to coding agents. The MemGPT research had demonstrated that LLMs could be given effective long-term memory through hierarchical memory management, compensating for the context window limitation that caused most AI tools to start each session fresh.

Letta Code applied this to software development: the agent maintained long-term semantic memory of codebase facts, working memory of the current task context, and episodic memory of past session history. On the first session with a new repository, Letta Code behaved like any other coding agent; by the third or fourth session, it had accumulated understanding of the project’s patterns, conventions, and architectural decisions — functioning more like a team member familiar with the codebase than a stateless tool.

At launch, Letta Code ranked first among model-agnostic open-source agents on Terminal-Bench, the benchmark for autonomous engineering task completion. The npm-based installation (npm install -g @letta-ai/letta-code) and support for Claude, GPT-4o, Gemini, and local models made it accessible without Letta’s proprietary server. The project was Apache 2.0 licensed, with v0.27.9 released on June 12, 2026 confirming active development.

Sources

Last verified June 12, 2026