Inspiration
We kept watching the same thing happen on memecoin launchpads. A meme takes off, and the winners are the bots that sniped the launch or the people holding supply across twenty wallets. Most of the people who actually liked the meme bought the top. We wanted to know whether a market could be designed so that having good taste in memes is what pays off.
What it does
Damper is a memecoin exchange on Solana. Each verified person trades from one wallet, and orders clear together in sealed batches at one price, so speed and extra wallets stop being an advantage. A learned market maker quotes into every batch and keeps the price closer to what the meme is actually worth. Each coin stays listed only while enough distinct people keep trading it.
How we built it
One market core in Python and PyTorch holds the batch auction, the ledger and the trading terms. The simulator, the PPO training and the FastAPI backend all call that same core, so the maker sees the same inputs in training and in production. The simulator runs thousands of coins at once on a GPU against a crowd of noise, retail and informed traders plus six scripted attackers. The on-chain side is an Anchor program: it ties each participant to one wallet using Solana Attestation Service badges on Token-2022, and its settlement ledger replays the core’s batches and matches every balance. A TypeScript service built with Hono and SQLite handles enrollment and sessions. A Jetstreamer pipeline pulls pump.fun history from the Old Faithful archive into ClickHouse. We clustered 724,686 launches into 3,289 memes and used how fast attention decays, a median half-life of about 12 minutes, to calibrate the simulated crowd.
Challenges we ran into
The hardest part was the mechanism design, not the code. There are a huge number of ways to build a market, and almost every rule that makes one safer also makes it more boring. We didn’t want a sanitized exchange. People trade memecoins because they’re fun, and we wanted to keep the parts that make them fun: getting in early and being right, a coin taking off because people really love it, cashing out when you want to.
So for every rule we had to ask whether it stopped the abuse or just stopped the fun. Lockups are a good example. Forcing people to hold would stop quick dumps, but a coin you can’t sell looks like a scam, so we dropped them. You can always sell on Damper. It just costs more right after you buy, and that cost is fixed when you buy so it can never be raised on you later.
The other hard question was how much power to give the market maker. It needs enough to step in when trading turns unhealthy, but a maker that holds coins and also decides which coins survive has an obvious conflict of interest. We ended up giving the agents room to tune the market inside hard limits, and leaving the question of which coins live or die to public rules that nobody controls.
Accomplishments that we're proud of
We built one market engine and reused it all across our codebase. The same engine runs our simulations, trains our agents and powers the live exchange, and what settles on Solana matches it exactlyt. Our market maker learns fast. It trains in about eight minutes on a single gaming GPU, which means we can rapidly iterate on our RL agents.
It also does well! We compared it with a market maker we wrote by hand, under the exact same rules, and the trained one priced memes more accurately 95% of the time. It also won back about a third of the trading activity that our safety rules had cost us.
The market got healthier on almost every measure we tracked. Ownership is one example: in our simulated market, the ten biggest holders owned about 7% of a coin, compared with about 16% on pump.fun’s curve.
What we learned
Memecoin markets don’t have to be this violent. Sniping, cornered supply and pump-and-dumps happen because the rules create a latency race and reward participants for hoarding supply and distributing it across many wallets. Once we took those rewards away, people in our simulator had no reason to do them. An inorganic pump that sent a coin flying on pump.fun’s curve barely moved it on Damper, and no small group ended up secretly owning a significant chunk of a coin’s supply.
Prices still move a lot, but they move with the meme’s genuine attention. A coin goes up when people start caring about it and comes back down when they stop. Our rules stopped abuse, but they also made the market slow and less fun to trade. The trained maker fixed that and brought traders back and kept prices honest, and it did both better than the maker we wrote by hand.
We also learned to look at where the money ends up. A smooth chart can hide ordinary traders losing to insiders, so we graded each market on who made money, who lost it and how much got stuck in dead coins. And we found we could test all of this in simulation before anyone put money in. We expect it to hold up with real traders.
What's next for Damper
Today the maker trains against six scripted attack strategies. Next, we want to train the attackers too: adversary agents that learn to extract as much as they can, with the maker trained against them, the same way game-playing agents get strong through self-play. Each round, the attackers find the weakest rule and the market learns to close it, so the rules are hardened before real money is on the line.
Then we’d open a live market with real traders and real money. Once Damper has its own order flow, the agents retrain on it, and we get to see how people trade with holding caps and exit fees when their own SOL is at stake.
https://damper-kappa.vercel.app/