SenseLab

SenseLab turns memory into a continuous learning system, where knowledge is shared, reinforced, and compounds with every task.

[Dai Macedo @daimacedoa]

We’ll be at AI Faire on Aug 28 with some of the other startups from our a16z Speedrun cohort :)

We’ll be showing what we’ve been building with SenseLab around memory + continual learning across agents, models and teams.

If you’re in SF, come say hi! We’ll have a little something for you too 👀

Invite only.

[Dai Macedo @daimacedoa]

3 weeks into a16z speedrun and it’s not at all what I expected.

I thought there would be a lot more “accelerator stuff.” sessions, frameworks, advice, etc.

instead, most of the time we’re just building. very fast 😂

I think the biggest surprise for me is the other founders. Someone is almost always a few steps ahead on a problem you’re trying to solve, has tried the thing you’re about to try, or knows someone you should talk to.

and the speedrun team is much more hands-on than I expected too. sometimes you don’t need advice, you just need the right intro or someone to tell you “yeah, this makes no sense” before you waste 2 weeks on it.

only 3 weeks in though. ask me again in October 😂

[Dai Macedo @daimacedoa]

I think I found a cheat code for founder-led GTM 👀

I put potential customers + everything I know about them into a SenseLab Room.

Now from Claude I can ask:
who’s a fit?
why?
what should I pitch?
write the outreach.

And it gets smarter as I use it.

Want access to this room? just comment ROOMS

[Dai Macedo @daimacedoa]

I might annoy a few people with this one 😂

But giant context windows are not memory.

Giving an agent access to everything that ever happened doesn’t mean it learned anything from it.

If it can’t tell what worked, what failed, and what should change next time, that’s just more context.

Not memory.

[Dai Macedo @daimacedoa]

Your agent shouldn’t have the memory of a goldfish. 🐟

We’ll be at AI Faire, Aug 28 in SF.

Come say hi, get some goldfish, and let’s give your agents memory that learns as they work.

@SenseLabAI + @speedrun

[Dai Macedo @daimacedoa]

basically the whole point of the sign 😂

regular memory = “I did this before”

learning = “I did this before, here’s what happened, and here’s what I should do differently next time”

[Dai Macedo @daimacedoa]

we call a lot of things AI memory
here's the difference :)

[Dai Macedo @daimacedoa]

If you want to try what we’re building at SenseLab and give me feedback:
https://t.co/XaCe4YyGhB

[Dai Macedo @daimacedoa]

I’ve been working so much that if I were my own employee, I’d probably sue myself 😂

“One more thing and I’m done” somehow turns into 11pm + 7 new tasks.

Anyway, back to violating my own labor laws.

[Dai Macedo @daimacedoa]

Claude learned it.

Cursor: who are you again?

switching AI tools shouldn’t mean starting from zero every time.

we’re fixing that with SenseLab.

[Dai Macedo @daimacedoa]

AI agents are great at explaining mistakes. We’re making sure they stop repeating them. 😅

[Dai Macedo @daimacedoa]

Big milestone for Bruno @balovespizza and me.
SenseLab is backed by a16z @speedrun

This is our third company together. We've spent years building infrastructure, and we started SenseLab because we found a problem we couldn't stop thinking about.

AI agents generate valuable knowledge every day, but almost none of it becomes shared across agents, teams, or workflows. We started SenseLab to change that.

Huge thanks to the entire a16z @speedrun team for believing in us. Excited for what's ahead.

If you're building AI agent solutions and you've run into challenges around shared knowledge between agents, I'd love to hear about it. DM me.

[Bruno Andrade @balovespizza]

AI came for our jobs, but it took a wrong turn when I checked the invoice 😂

We’ve been testing this at @SenseLabAI

On CI fixes, a specialized model trained from an agent’s experience solved 94.2% of tasks vs 78.7% for the frontier model without SenseLab.

At 1/16 the model cost! 💰

Then we used @SenseLabAI to train a model on the previous model’s own traces produced using our Recursive Self-Improvement engine. Its first-try results improved too.

That’s what gets me excited about recursive self-improvement.

Imagine your support queue or recurring engineering tasks producing the experience to train your next model.

Your expensive AI gets a new assignment: train the cheaper, expert hire. 🤓

[Bruno Andrade @balovespizza]

How the third one works: act → see outcome → rewrite what it believes. A fix that works gets trusted more. A trusted fix that fails once on an unsolved task is demoted on the spot. When its own judgement runs out, it works through untried options in order instead of guessing — new fix found in 1–2 tickets, not 30.

Same model (gpt-5.4-mini), same prompt, same 3-attempt budget. CI queue, same setup: 78 → 76 → 8.

The engine is open source: https://t.co/8Lwf4Hff0M

[Bruno Andrade @balovespizza]

We gave the same model its entire history in the prompt. It barely helped.

Same 60-ticket queue, three ways. Rules change twice mid-run.

fresh context: 88 wasted attempts whole history in prompt: 77 learns from its own outcomes: 26

Context isn't learning. This is what learning looks like.

[Bruno Andrade @balovespizza]

Somehow, people are just accepting agents delivering mistakes as "normal", and so many products are being built on top of it.

Teams have the hope that the model alone will one day improve enough to get it all right, or that just "one more prompt or md tweak will fix it"

The truth is, it won't, because your agents need to self-improve based on your data and what they learn from your environment.

We are fixing that at @SenseLabAI

[SenseLab @SenseLabAI]

@hayesdev_ The same applies to debugging: most time isn’t spent fixing, but rebuilding context.

That’s what we’re exploring with @SenseLabAI

[SenseLab @SenseLabAI]

@PeladoNerd I loved the "brazos y ojo" definition, its exactly what we are! Thank you

[Pablo Fredrikson @PeladoNerd]

Un agente AI es como agregarle brazos y ojos a un LLM (🧠)

Actúa como un compañero autónomo que puede encargarse de tareas operativas por vos: investigar alertas, responder a incidentes o incluso ejecutar consultas contra tu infraestructura.

#ad