Letta Code: A Memory-First Coding Agent Fresh

Published: December 16, 2025 | Original Post

The #1 model-agnostic open source agent on Terminal-Bench.

Key Features

Continual Learning and Memory

Three mechanisms for accumulating and applying experience:

Memory Initialization -- The /init command triggers agents to conduct deep research on existing codebases, forming memories and updating system prompts via memory blocks
Skill Learning -- Agents transform repeated task patterns into reusable skills (database migrations, dashboard creation, API best practices). Skills are stored as markdown files for version control
Persisted State -- Agents search through past conversations using vector, full-text, and hybrid search via the Letta API

Model Flexibility

Letta Code adds statefulness and learning to coding agents while maintaining model flexibility. It achieves performance comparable to provider-specific tools like Claude Code and Gemini CLI.

Performance

According to Terminal-Bench results, Letta Code is the #1 model-agnostic, OSS harness in comparative testing.

Getting Started

bash

npm install -g @letta-ai/letta-code

Documentation available at docs.letta.com.

Letta Code: A Memory-First Coding Agent Fresh ​

Key Features ​

Continual Learning and Memory ​

Model Flexibility ​