Letta AI · 2026-05-04 · notable
Letta Code v0.25 — Memory-First Coding Agent, Persistent Context Across Model Switches
Open-source coding agent (Apache 2.0) that persists memory across sessions and model switches. Memory subagents review sessions and rewrite context. Works with Claude, GPT, Gemini, Ollama, or any provider — agent identity is decoupled from the model.
Letta Code stores coding context in persistent memory — switch models mid-session and your agent remembers everything.
Key specs
| GitHub stars | 2,424 |
|---|
What is it?
Letta Code is an Apache 2.0 open-source coding harness from Letta AI (formerly MemGPT) that launched April 6, 2026. Unlike Claude Code or Codex, the agent's identity and accumulated knowledge are stored in a persistent memory layer separate from any specific model.
How does it work?
Agent state is stored in Letta's memory architecture: as you work, memory subagents periodically review session transcripts, extract durable knowledge, and rewrite the agent's context store. This means switching from Claude to GPT or Ollama mid-project doesn't reset the agent's understanding of the codebase. Available as a CLI, desktop app (Mac/Windows/Linux), and MCP-compatible server, with skills-based extension support.
Why does it matter?
Most coding agents are stateless — every session starts cold. Letta Code's persistent memory means agents accumulate project-specific knowledge over time: coding conventions, architecture decisions, known footguns. This compounds across sessions rather than resetting.
Who is it for?
Developers who work on long-running projects and want an agent that genuinely learns their codebase, or teams wanting model-agnostic agent infrastructure.
Try it
npm install -g letta-code && letta-code