I said “ugh” out loud when I saw the results this morning. NMA didn’t win. The Qwen Cloud Global AI Hackathon — 729 submissions, five tracks — and in the MemoryAgent track we’d entered, the winner was someone else’s project. In Build Notes II I wrote that the eleven days were the most important ones. Back then I still believed the results would finish that sentence for me. They didn’t.
My reasons for being disappointed were embarrassingly simple: no money. The track winner took home $7,000 in cash; the honorable mentions got $500; I got neither. And I’ve been poor lately, lol.
But honestly, the outcome was always a lottery. Five tracks, seven hundred submissions, judges are human, and judgment is subjective. I also never played the competition game — aren’t Chinese people supposed to be the best at test-taking? I did the opposite: I built the whole thing behind closed doors. What I submitted was a system I genuinely use while writing my novels. It has flaws, and plenty of it came from my own imagination. But it really did help me keep my characters straight, and it really did run. So by my own accounting, I won. That’s enough.
Then I went and read the winners’ code.
Of the five projects that won in the memory direction, three open-sourced everything — Quên, Reverie, Engram. The other two at least wrote their submissions up properly. I handed the who-won research to Cask, my AI assistant — the kind who arrives with things you didn’t ask for. He came back with more than a list: all three repos pulled, read down to the implementation, the other two turned inside out, the whole thing organized into a research doc for the next project. The more I read, the more absurd it felt. Not “how did they win” absurd. “They won, and I can just learn from them” absurd. Judge-approved solutions, handed to me to shear. At that moment I wasn’t disappointed anymore — I was getting wool, and even the no-money regret was gone.
The winner first. Quên — Vietnamese for “to forget,” and it sounds like Qwen — is a memory system that knows how to forget. It computes a forgetting curve for every memory using FSRS spaced-repetition math. Low-confidence memories get verified against the source before the system answers, and the verification feeds back into retention: confirmed memories get stronger, refuted ones get discarded on the spot. Its entire evaluation cost $1.6. Its sharpest line lives in the README: trust is a runtime decision, not a stored property.
The others earned their place in my notes too. Reverie has a rule called “No quote, no memory”: a memory can’t enter the store without the exact conversation quote as evidence, and the model can propose changes, but only deterministic code can commit them. It also keeps a memory type called affect — not a dry record of “she was anxious,” but “here’s what helps when she’s anxious.” Engram turns forgetting into graduation: superseded facts retire instead of being deleted. Qwen Memory Agent pins user preferences so they never decay. NaLog runs in Thai rice paddies, and every action is propose → approve — the model never has permission to switch on the pump itself.
Put the five together and it’s one consensus: forget explicitly, verify, and be able to defend it. None of them won by remembering more. They won by knowing what to forget, what to trust, and where each memory came from.
And those are exactly the problems I was circling with NMA — forgetting, confidence, sleep consolidation. The direction was right. Other people just maxed out the counterintuitive part: forgetting.
In The Scaffold of Time I wrote that everything about the desktop AI companion I’m designing is still an idea. Now the idea has its first batch of ammunition: five judge-approved projects and one industry consensus. I sorted them into three piles — what goes straight into the memory architecture’s contract, what’s reference-only, and what I need to think about before deciding.
I studied other people’s work, and then I started building. 😂