We're releasing Letta Code, a memory-first coding agent - open source (apache 2.0) - model agnostic - portable agent learning and memory

Letta Code is also the #1 model-agnostic OSS harness on @terminalbench - with comparable performance to model-specific harnesses (Claude Code, Gemini CLI, Codex) across frontier models

GitHub: Blog post:

try wargrep

Letta Code 🤝 WarpGrep Subagent?

yessir but your dm's are closed

Very very cool! Would love to chat with you guys

why would you make it look and feel like claude code? this is not allowed. claude code has actual fanboys

Huuuuuge

Holy heavens!

Memory-first agents are the future.

Impressive benchmarks!

Whooooo!!! Go @sarahwooders @charlespacker

super interesting! Any plan to add ACP? I couldn't find any reference in the documentation.

Super interesting, congrats! Will give it a try

how is this different from claude's existing context management stuff?

benchmarks or nothing happened

Very nice 🙌

This project + my energy = unstoppable! 💎💯

Memory-first is the right approach. Stateless agents forget context and repeat mistakes.

Omg, a coding agent that remembers? Finally, my code won’t forget me 😂

can i get the pro for free for a month to try?

Free plans include usage for "Auto" and base tier models, and you can also bring your own keys or external subscriptions!

Memory-first is the right framing. Most coding agents treat context as disposable, then wonder why they lose track of goals mid-task. Portable agent learning is wild when you think about how much gets lost between sessions today.

Awesome

What is memory first

This is basically what I was trying to build as a smaller side project. Very interested to try this out

Free coding credits to kickstart as advertised:

This is awesome! Congrats on the launch—looking forward to seeing what people build with it.

Welcome! You will get the 5 min test as every new AI tool deserves nowadays. 🙂

awesome, documentation is outstanding as well

I've been using Leeta for a while. While it can remember things, it feeds all the memories to the AI ​​model to generate answers every time, causing token costs to increase exponentially. I've already had Leeta refine the memories, but the results are still unsatisfactory. Moreover, at this rate, the cost will continue to double with each memory. So I'd like to know if the team is continuing to optimize this aspect of the product?

@theCTO the logo haha

no windows support 🥲

amazing! how does it compare on benchmarks against opencode?

hmmmm

Nice will try it out

Congrats! If running through docker when configuring models can we configure base urls per model?

Is BYOK supported?

Yes! Two ways to do it - (1 / easy) set up BYOK on your Letta Developer Platform account (inside account settings), (2 / more advanced) self-host the Letta server itself, and provide your LLM API keys when you start the Letta server (

Looks great! I’ll try it out on Windows to see how it performs. Many CLI tools for coding don't work that well on Windows (they often default to bash, and not familiar with PowerShell 7 syntax).

Let us know how it goes! If there are any Windows issues, happy to debug, just open an issue on GitHub. The defaulting to wrong non-windows syntax issue does sound like something the Letta Code memory / learning can solve

It failed at the first step.

memory-first coding agent that learns and persists across sessions is the architecture that actually changes production value. stateless agents feel dumb because they forget everything. portable agent learning turns one-time interactions into a compounding knowledge asset.