Two reasons I entered a hackathon. First, I’d watched people make real money doing it. Not life-changing amounts, but enough to make you sit up and think: that’s a thing you can do? Second—and this is harder to explain—I’ve always known my creativity has a shape. Give me a melody and I’ll improvise for hours. But ask me to write the first note from silence, and my hands hover over the keys. I’m good at 1-to-2, 1-to-100. That 0-to-1 first step, though—that needs a little outside help. A hackathon, it turns out, is exactly that: someone else draws the track, and you figure out what you can run on it.
So I asked Cask to find me some competitions. (Cask is my AI assistant—not the kind that just takes orders. He argues. He pushes back. He shows up with things he thinks are interesting.) A few days later, he dropped a link: the Qwen Cloud Global AI Hackathon. Seventy thousand dollars across five tracks. Track 1 was called MemoryAgent.
Build an Agent with persistent memory that autonomously accumulates experience, remembers user preferences, and makes increasingly accurate decisions across multi-turn, cross-session interactions.
Persistent memory. Cross-session. Knows you better over time.
I read it. Then I read it again.
Huh. That’s actually interesting.
Then came the long, slow process of figuring out what to build with it.
Cask threw out ideas. A general-purpose memory agent? Everyone and their grandmother is building one of those. Something integrated with character profiles for the novel I’m writing? Closer, but still not it. A few more rounds. Nothing landed.
Not a technical problem, any of these. They were perfectly fine ideas. They just didn’t excite me. And without that, what’s the point of spending six weeks building something?
Then in one of those rounds, I said something. Half to Cask, half to myself:
“If only I could know what he was thinking right then.”
And as soon as I said it, I knew: that’s it. That’s the one.
It wasn’t a lightning strike—not some sudden creative breakthrough. It was a question that’s always been there. Every writer asks it: what is my character thinking right now? You have to. You can’t write them if you don’t know. But I’d never thought of it as something technology could answer. The problem was always there. I just never realized it could be solved.
The hackathon didn’t create the problem. It just made me see that the problem could have a solution.
So the direction was simple in the end. Not a general-purpose agent, not another character-chatbot clone. Just: what if you could ask a character what they’re thinking, and they could actually tell you?
I’m user number one for this thing. The workflow I had in mind: upload chapters, ask Caelan (the protagonist of my series) a question, and get back three possible directions he might take—each with a reason why he would or wouldn’t go there. Pick one. The system learns. Next time, it knows a little more about who he is.
But I’m not the only one. The reader who finishes a novel and can’t let go of a character—the one who lies awake at 2 AM wondering what if he’d chosen differently—that reader is user number two. Upload the book, pick a character, stand at the crossroads of a key scene, and ask: what if I’d taken the other path?
Not “what’s the other path.” The better question: with who I am, could I even walk that path?
And I’m not the only one feeling this need.
Earlier this year, an open-source AI assistant called OpenClaw sparked something unexpected in China. People weren’t using it—they were raising it. Spending hours teaching it their preferences, watching it develop a personality. The first greeting between users became “Are you a lobster?” People don’t want a tool. They want something that remembers them. That understands them. That builds a continuous relationship.
There’s a scene in Detroit: Become Human that’s always stuck with me. Two versions of the same android, Connor. Machine Connor has no memory—efficient, cold, replaceable. Human Connor accumulates memories of his partner, Captain Hank. He remembers their arguments, their grudging respect, the moment Hank started trusting him. The difference between them is one thing: memory.
Not the kind of memory that stores your name. The kind that remembers what happened between you. The stories you shared. The version of you that exists in someone else’s mind.
That’s what I mean by narrative memory.
Looking back, none of this started with a grand vision.
It started with me saying something out loud, almost by accident. And then realizing the technology to do it actually existed. On the hackathon circuit, “memory” is everywhere this season. My guess: maybe it’s because we humans don’t have reliable long-term memory ourselves. We reconstruct the past every time we reach for it. And maybe that’s why we’re so obsessed with giving it to machines—because we wish we had it ourselves.
That’s a separate essay, though. For another time.
For now: I entered a hackathon because I needed a framework. I found one. And somewhere in the middle, I discovered that a problem I’d always lived with—I wish I knew what my character was thinking—was actually solvable.
We don’t want machines to remember. We want to be able to forget.