13/05/2026
Building... a brain for an AI agent.
This is going to be an interesting story, so if you've got a minute — welcome to my (our) world.
Quick honest note before I start: the current pace of growth and skill development in our team is something I used to dream about. If someone told me two years ago that we'd be building what we're building today, I honestly wouldn't have believed it. AI gave us near-exponential growth in capabilities and speed of shipping — solutions I could only dream of just a year ago.
At the same time — to be fully honest — the last few years have brought a record number of failures (more than the rest of my life combined), but also a record number of attempts. That definitely fueled our pace. I also feel the gap between our potential and our revenue is the widest it's ever been — but we're working on it 😃
Anyway — back to the story. At a recent Python developer conference, Borowiecki gave a talk on Cognitive Context Engineering. Genuinely solid talk — thanks Olgierd! That said, I left feeling there wasn't really anything in it "for me" at that moment, talked to a few people, and called it a night.
I woke up at 2:30 AM with my brain on fire, full of ideas and solutions. Ever since I started taking my sleep and circadian rhythm seriously — and dropped alcohol — I get these occasional nighttime "peak flow bursts": I wake up with an idea AND the full architecture of a solution already in my head. The previous "burst" turned into a product that's now a registered EU trademark, and we're launching its promo soon (that's a separate post XD).
This time I woke up convinced I had to solve a specific set of problems around context and memory for AI agents — so that our main agent could:
— have both short-term and long-term memory,
— understand the broader context: what we're doing, who we're doing it for, the roles in the team (both AI agents and humans),
— know what it's DOING WRONG, what to learn, and what to unlearn,
— create things autonomously — either planned, or on our commands,
— know which external tools it can use,
— know the hierarchy and what it can share with whom at given permission levels.
These are obviously high-level functionality descriptions — I'm not going into technical detail in this post.
Now, some time back I'd glimpsed this "neural memory graph" thing but couldn't pin down which solution it actually was. I'd heard of Obsidian, but it wasn't enough to build something bigger and more scalable. That night, everything clicked into place in a single moment — I found Neo4j (with LangGraph), connected it with our other tools and processes, and designed the architecture.
Mati — our AI vice CEO — is being built right now, and early results are very promising ;)
Every time I tell this story or share examples of other solutions we've built, friends keep asking: "Why aren't you posting about this on social?!" For a long time, while doing genuinely good work for other companies — real revenue, real clients — I was stuck on a well-known cognitive bias that kept me from publishing. That's changing now.
If you're interested in what comes next and the systems we're building for our clients, follow our two pages: Matt Jey and Brandesun. Talk soon!
P.S. I wrote this one myself — Claude only had a peek 😄
