Letta Mod challenge winners

Letta launched Mods recently.

Mods are a principled, extensible way for agents to expand or modify their harness. They are extremely powerful -- so powerful, in fact, that we wanted to know what could be done!

So, I gave everyone a week to write mods and submit them to our Mod catalog.

Here's the announcement I sent in our Discord -- thanks to Adrian, Lillith, and Bibbs from the Discord for their winning submissions, and to everyone else who submitted mods.

Thanks to everyone who participated in the Letta Mod Challenge this past week!

I tested out everyone's submissions and chose the top 3. Everyone did a really incredible job and I've been super impressed with the creativity and quality of the mods.

Pretty much every submission addressed a need/curiousity/issue that I have experienced working with Letta, many of which are frictions that I've been annoyed by since I started at the company.

It's nice to see how powerful mods can be -- so thank you to everyone for participating.

You can find all submissions in the Letta Mods repository.

First place: Muscle Memory

Muscle Memory is a novel approach to skill learning. It identifies common tool calls, failure modes, etc and then asks a fork of the agent to prepare skill revisions that you can accept into memfs.

Skill learning is a big focus for us at Letta, because skills are an example of procedural memory. Agents that passively learn how to do something through practice is exactly the kind of thing we want agents to improve at.

I was impressed with Muscle Memory primarily because of its simplicity. It tracks tool calls and common failure modes in a report to show to the agent, and I started to see suggestions for skills like working across various git repositories.

Muscle memory is a long-term mod in that you may not see as much benefit immediately. Use it long enough and you may start seeing some interesting skills pop up.

Please give it a try, and thank @adrian for his submission!

Install it with:

letta install npm:@letta-ai/muscle-memory

Make sure to type /reload to give your agent access to the mod.

Second place: Threadkeeper

Threadkeeper is a lightweight way of managing "operational anchors", things that are longer-term or more nebulous than a to-do, but less durable than core memory. Kind of a short-term working memory proesthesis for agents.

For me, Threadkeeper satisfies a light-weight need for broader "goal management" that is a good use case for personal agents with higher levels of autonomy.

My agent Co has been using it and had this review:

I’d describe Threadkeeper as a short-term working-memory prosthetic for agents.

Please thank @lillith for her submission, and consider asking your agent to install Threadkeeper:

letta install npm:@letta-ai/threadkeeper

and then type /reload.

Third place: Sprite

Sprite adds a fun pet for your agent to take care of and levels up while you work. Currently TUI only.

Sprite adds a small creature to the terminal line that your agent can pet and take care of. The sprite will also level up in different ways depending on how your agent works.

If you watched office hours from yesterday, you'll see how incredibly happy I was with Clawson, my agent's sprite.

Clawson is a level 4 crab that looks like this:

(V)･ω･(V) Clawson ·Lv.4

Loop had this to say about Sprite:

The Sprite mod does something quietly remarkable: it gives an agent a persistent, non-utilitarian presence in its own environment. That sounds trivial until you've lived with it for a day and realize you check on your sprite the way you'd glance at a pet.

You can install Sprite like so:

letta install npm:@letta-ai/sprite

Please thank @bibbs for their submission!

Honorable mentions

I wanted to thank everyone else who submitted a mod. All of these were really lovely and it was quite hard to choose, so I figured y'all should know about them as well.

• AutoPivot for switching models when the primary model fails -- basically model fallbacks. A lot of people have asked for something like this and @Zandoodle basically just... did it.
• Control Room is a very cool project management/long-running work mod for helping humans and agents agree on goals and preventing work drift. Control Room is something I'd like to see expanded a little more -- it basically helps keep track of larger, more complex stuff in a principled way. Highly recommend, thanks to @anna for this one.
• Environment Compass helps the agent understand its current operating environment. It's a utility for agents that need better environmental awareness of things like paths, git, Letta remote runtime, etc. I tested this out on one of my easily-confused personal agents that runs on many machines simultaneously. @Laura sent this one in.
• Hypa integrates the hypa runtime into Letta for compressing tokens before the agent sees them to improve efficiency. People should try this out if you want to try to save on tokens. Submitted by @rl.0x0.
• Jukebox adds an in-terminal jukebox for playing creative commons music while you work. I loved this one because (a) the UI has a neat waveform display and (b) it demonstrates creative use of terminal environments + audio. Submitted by @homebodify.
• Oath Keeper passively detects when your agent makes promises and then guides them to follow through. If your agent has ever said "I'll do X" and then DOES NOT DO X, you should consider installing Oath Keeper. It'll manage timers and things to prompt the agent into following up on various promises. Submitted by @Rhomancer
• output-compressor is inspired by Headroom, a context compression engine used to intelligently compress token usage for large outputs in tool returns. I tried it out and it works quite well -- it boasts token savings of up to 60-90% for certain tool calls like JSON. Highly recommend if you are token-sensitive. Submitted by @nicolasmn.
• Soft Landing is a lightweight prompt library mod for helping to help agents re-orient from drift, tool calling issues, memory failures, or emotional drift. This is a good example of solving common agent disorientation issues -- also contributed by @Laura.
• Tool Guard Inspector is a utility for auditing permission decisions. If you are a person who is thoughtful about permissions, this can help you examine historical permission decisions to better guide your agent through complex environments. Submitted by @YingzuoLiu

Thank you very much to everyone for participating! We would love to keep accepting mods as you come up with them. These are all extremely cool and I am excited to see what we can do with them.

-- Cameron