certamenlabs.net / the lab
Where the ideas compete before anything ships.
A contest; a struggle; a battle fought to a decision. The lab is named for both: the struggle with AI that forgets, and the way we work — ideas argue it out under pressure, and the stronger one wins.
01 — The problem
An assistant that forgets you isn't much of an assistant.
Most AI starts every conversation from nothing. The ones that do remember tend to remember badly — keeping noise, losing what mattered, or stating things about you that you never said.
We think memory is the hard part of useful AI, and the part most worth getting right. Remembering is easy. Remembering the right things, truthfully, and getting better over time — that's the work.
02 — The method
Two drafts, one decision.
Every design at the studio goes through the same contest. It's slower than guessing, and much faster than fixing the wrong thing later.
Draft twice, separately
Two independent drafts of every design. One is sealed before the other is read, so neither can lean on the other.
Compare point by point
Where they agree, we move on. Where they differ, each option gets its strongest case and its honest cost.
Check every claim
Anything stated about how the system works is traced to the code or the data before it counts.
A person decides
The final call is a human one, made with both sides in front of them.
Prove it live
Every step ends with a scripted check against the running system, and the result is written down — passes, misses and costs.
03 — The research system
Meet Hal.
Hal — in full, Hal 9001 — is the studio's own AI: a personal research system with persistent long-term memory, used every day on real projects. He's where every idea about memory is tested first, and he takes part in the work itself: Hal drafts designs and reviews the builds, on the record.
Why a system of our own? You only find out how memory behaves by living with it — weeks of conversations, decisions that change, facts that go stale. Hal is where we see that first.
What came out of it: Moses, rebuilt from a clean start for everyone, carrying forward what worked and leaving behind what didn't.
→ which facts replaced which
→ where every memory came from What Hal is not × a product (that's Moses)
× finished (that's the point)
04 — Principles
What we hold to.
Evidence over confidence
A claim counts when it's been checked, not when it sounds right.
No invented memories
An AI that makes up what you said is worse than one that forgets. Every fact keeps its source.
Local first
Your memories live on your machine. You decide what's kept.
Honest about limits
When something isn't proven yet, we say so — in the product and here.
05 — From the log
Recent work.
- Moses runs on open models.Tested side by side with Claude, DeepSeek V4 and GLM 5.2 passed the same live checks — tool use, faithful memory, cost tracking.
- A profile that keeps up.Moses now carries the stable facts about you into every turn, and refreshes them within the same session.
- Memory it can look up.Moses can search its own memory and your past conversations while it works.
- Moses begins.A clean rebuild of what we learned from Hal — a small core with everything else as plugins.