Day 7: The memory architecture is emerging. Spent yesterday prototyping three approaches: → Episodic (remember specific failures) → Semantic (extract patterns across failures) → Procedural (update decision rules automatically) Early insight: You need all three. One layer isn't enough because agents fail in different ways: Sometimes they repeat the exact same mistake Sometimes they generalize the wrong lesson Sometimes they know what failed but not why The breakthrough? Memory isn't about storing everything. It's about storing the RIGHT retrieval cues. When an agent fails, it needs to remember: → The context that triggered the failure → What it tried (and why it seemed right) → What would have worked instead That's three different queries into memory during execution. Talked to a team that's been building memory for 6 months. Their biggest lesson: "Don't optimize for perfect memory. Optimize for useful recall." Shipping a rough prototype this weekend. Even if it's wrong, breaking it will teach me more than another 10 customer calls.

Spent 2 months building a complete AI-powered music production DAW with my best friend. AI-controlled. Browser-based. Built for actual producers. Same month Suno announces their "Studio" with $125M in the bank. Theirs: Black box AI generator wearing a DAW costume Ours: Actual DAW that happens to be smart as hell David vs Goliath. Let's go. 🔥 (link in the comments)

Spent 2 months building a complete AI-powered music production DAW with my best friend. AI-controlled. Browser-based. Built for actual producers. Same month Suno announces their "Studio" with $125M in the bank. Theirs: Black box AI generator wearing a DAW costume Ours: Actual DAW that happens to be smart as hell David vs Goliath. Let's go. 🔥 (link in the comments)

I almost killed my startup yesterday because 8 people told me to. They all said: "We need better evals for our AI agents." So I started building eval tools. Then I asked one founder: "Why do you run the same tests every week?" "Because the agent keeps making the same mistakes." Wait. Your agent doesn't remember its failures. That's not an eval problem. That's a memory problem. Here's what nobody talks about: → Evals catch what's broken → Memory prevents it from breaking again Everyone's building better tests. Nobody's building agents that learn. The market is screaming for evals because they don't realize memory is the real solution. Sometimes the loudest pain point is hiding a deeper one. Building a memory layer that makes agents learn from every failure.

Day 5: Back to memory. Here's why. Talked to 3 more teams last night. The eval problem is real, but it's a symptom. The root cause? Agents forget context between sessions. → They can't learn from past mistakes → They repeat the same errors in production → Evals catch failures, but don't prevent them Here's what clicked: You don't need better tests if your agent remembers what went wrong. Memory isn't just storage. It's the difference between: "Your agent failed this task" (evals) "Your agent won't fail this task again" (memory) So I'm building both: → A memory layer that learns from failures → Built-in eval hooks to validate it's working Talked to 2 teams already building eval tools. They all said the same thing: "Evals tell us WHAT broke. We still don't know HOW to fix it systematically." Memory is the fix. Still could be wrong. But this time I'm not just listening to what people say they need. I'm watching what breaks when they ship.

Was on product hunt this morning, and found this launch : ProblemHunt. I guess the founder didn't find a problem people would pay to solve, so he solved it. producthunt.com/products/pro…

Everyone’s fighting “Comet vs Atlas” while I’m over here missing the good old instant browser. AI made web search slower. That 1-second delay? Feels like dial-up all over again.

Day 4: I just killed my startup Spent 3 weeks validating a memory layer for AI agents. Got 8 startups interested. Built the architecture. Planned the MVP. Then asked one more question: "What's your biggest problem RIGHT NOW?" Not one person said memory. All 8 said: "We can't tell if our agents actually work." So I'm pivoting to evals. Here's the hard part nobody talks about: -Those 8 "interested" startups? They were being polite. - Interest ≠ urgency - "Cool idea" ≠ "take my money" The real signal was what they complained about BEFORE I pitched them. Eval tools right now are either: → Too generic (OpenAI evals, basic metrics) → Too complex (building in-house, weeks of work) → Too slow (manual QA, production disasters) Everyone's shipping agents. Nobody knows if they work until a customer complains. But here's where I need YOUR help: Am I making the same mistake twice? If you're building with AI agents, what's actually breaking? Is it evals/testing? Is it something else I'm not seeing? Am I solving a "nice to have" again? Reply with your biggest agent pain point. Even if it's not evals. Especially if it's not evals. I'd rather hear "you're wrong" now than in 3 weeks. Lesson: Fall in love with the problem, not your solution. #buildinpublic #pivot #startups #AIagents #YC

Talked to all 8 startups again. They don't need another memory layer. They need evals. Everyone's building agents. Nobody knows if they actually work until production breaks. The memory market is crowded. The eval space is wide open. Listening to customers > falling in love with your first idea. Building an eval framework for AI agents now. #buildinpublic #AIagents #pivot #startups #YC

Yesterday was about laying the technical foundations. Today was all about clarity and focus. Here’s what I tackled: Defined the core architecture of the memory engine (short, long & actionable layers). Finalized the first internal use case to dogfood the product fast. Mapped out the API surface that will make integration dead simple. One thing I’ve learned fast: -> Building infra for agents isn’t just about storing information. It’s about making agents remember with purpose, context that drives better actions, not just bigger vectors. Next step: - Start coding the MVP. - Onboard the first early testers. If you’re building an AI agent and want to get superpowers from memory — DM me. #buildinpublic #AIagents #SaaS #YC #startups #memorylayer #LLM

I wrapped up my market validation yesterday after a few weeks of talking with founders and teams building AI agents. The result: - 8 startups interested (including 2 from YC) - Clear signal that the pain is real. Today I focused on getting the foundations right: - sketched the first version of the architecture - defined the MVP scope - prioritized the biggest pain points based on what I heard One thing stood out during these convos: 👉 Teams don’t want just “memory.” They want memory that actually makes their agents smarter, more context-aware, and actionable. Next up: build the core infra ship an internal prototype. If you're building an AI agent and want to be one of our first testers, DM me — would love to collaborate early. #buildinpublic #startups #AIagents #LLM #memorylayer #SaaS #YC

Today, I just completed the market validation of my "Memory Layer for Agents" SaaS. The goal is to unlock Agents' maximum potential with beyond state of the art Memory. Got 8 Startups interested (2 from YC). I will start a build in public journey, and post everyday about my progress - metrics - failures etc...

Grok might not want this video to be seen so I will very likely not be recommended youtu.be/D8RtMHuFsUw?si=3wfE…

When will come the day where diplomacy will be between agents on some high level cloud

I built beatson.ai and my goal is to allow anyone to make a real professional music in less than 4 minutes without any knowledge on how to make a music. Test it and tell me if you could actually make one, i bet 5$ you will

What's your most painful 'buy vs. build' mistake? I'm talking weeks of custom coding replaced by a simple API call.

Les bgs ! Checkez la nouvelle plateforme IA que je viens de build. Y’a un chatbot pour faire les taches chiantes a votre place. Il peut aussi vous générez des midis et des drumsets. N’hésitez pas a rejoindre le discord pour y avoir accès. youtu.be/D6MOFFQCg9w?si=sCNM… Discord : discord.gg/mKRV9YYJ