Your AI Doesn't Forget. That's the Problem.

An agent that turns one bad trade into a permanent rule is worse than no agent at all. That's the whole argument in this week's AI section: a skill that teaches your setup to log a mistake as a data point, not a policy, and to flag the moment you're about to contradict a decision you made three weeks ago.

Turns out the market needed the same lesson this week.

Robinhood now earns more from prediction markets than from stock trading, and it feels like it might be the case for the trenches as well. CEOs are already following memecoins paired to their own stock, which is either an anecdote or the start of a pattern, and we won't know which for a while yet.

What we do know is when we'll get the first real read. CPI lands Thursday, FOMC comes the week after, and everything else this month is either setting the tone for those four days or quietly landing in their shadow. If you're carrying size into either date, that's a decision made this week, not on the morning it happens.

As for the Miles High Club update: Surfxbt is printing on his HYPE trade. Paradise built a second position on a fresh trigger instead of holding last week's winner out of hope. Morin picked one conviction call and sized everything else around whether it fires. And much more.

Let's dive in.

Last week you gave your agent your rules. This week, teach it which ones to keep.

We walked through the context file: risk limits, sizing, invalidation levels, watchlist. One .md file that turns a generic model into something that reads the market the way you do.

Here's the part nobody warns you about. Once your agent starts remembering things across sessions, it remembers the wrong things. It keeps a throwaway comment you made at 2am and treats it as policy. It holds onto a thesis you abandoned three weeks ago. Or it forgets your sizing rules entirely and cheerfully suggests a position twice the size you'd ever take.

A skill published this week fixes exactly that, and it's the most underrated thing in AI right now because it's boring. It's called agent memory discipline, and it teaches the agent when to remember rather than what.

One memory, one durable fact. With what, why, when, and where it came from. Not a transcript of the conversation.

One failed run is not a permanent policy. This is the one that matters for traders. You got stopped out on a breakout long. That's a data point. It is not a rule that says breakouts don't work, and an agent that quietly converts one bad outcome into a permanent bias is worse than no agent at all.

No credentials, no personal data. Obvious, routinely ignored.

Don't recall on trivial questions. If every message triggers a memory lookup, you've built latency and noise, not intelligence.

When a newer decision replaces an older one, the skill closes the old memory instead of deleting it. Then it surfaces both, and flags the contradiction rather than silently picking a side.

Sit with that for a second. You now have an agent that can tell you: three weeks ago you said you'd stop trading majors below the 200-day, and today you're asking about a long below the 200-day. That's not a memory system. That's the trading journal you've been meaning to keep since 2021, except it reads itself back to you at the moment you're about to break your own rule.

The skill lives at buildwithclaude.com/skill/agent-memory-discipline, and the setup prompt handles it end to end: it inspects what you already have, installs at project or global scope, asks what memory backend you're using, then defines the policy.

The step people skip is the proof run. Save one real fact, an actual convention or decision from your own trading, then open a fresh session and check what comes back. If the recall is wrong or noisy on one fact, it'll be wrong on a hundred.

20 minutes. It's the difference between an agent that knows your rules and one that has opinions about them.

A recurring slot for a skill or prompt you can install this week. No essay, just the paste-in.

This is the skill behind the AI x Trading section above. If you read that and want the actual setup rather than the explanation, it's below. If you'd rather skip straight to using it, copy the prompt and go.

Deliberate recall. Durable saves only. One memory, one fact. Superseded history kept, not deleted. Contradictions surfaced, not silently resolved.

Last week we showed you the activity numbers: Pump.fun's mobile app past $50M a day, Robinhood Chain through $900M in a session, one relay pulling $640K in fees in twenty-four hours.

The obvious question was what happens when it cools. Four things this week suggest the more useful question is what it turns into.

Robinhood now earns more from prediction markets on its platform than from people trading stocks.

Look at the actual line, and it's the "other transaction-based revenue" bucket, which is where event contracts sit, so treat it as directional rather than a clean single-product number. But the direction is not subtle. That line crossed above equities revenue last quarter, roughly $205M against about $129M, after years of sitting flat near zero while equities did the work.

A brokerage built on commission-free stock trading now makes more money from users betting on outcomes than from users buying shares. That's not a crypto story and it's not a fad quarter. It's a business model reorganising itself around speculation as the product.

Rune caught the part that doesn't show up in any dashboard.

Vlad Tenev had a public back-and-forth about tokenized stocks with AMC's CEO, then followed $MEME, a memecoin paired with tokenized AMC. Andrew Dudum, CEO of HIMS, followed $BONER, a memecoin paired with his own stock. Dudum follows eight accounts.

Read that again. Sitting CEOs are engaging with memecoins paired to their own equity. Every previous attempt to sell tokenized stocks to institutions went through conferences, pilot programmes and compliance decks. This one is going through memecoins, and it's working faster.

That's the mechanism worth understanding: the trenches aren't competing with tokenized equities, they're distributing them.

Kyle's framing is the cleanest we've seen, and it lines up with the revenue data.

Phase one, in his read, is done. Memes have no lasting power beyond the top few, grifters arrive to drain liquidity, it spreads to other chains, liquidity gets thin, it breaks. Phase two is where people start hunting for signal to differentiate, and capital drifts toward things with fundamentals. His three: AI, perps, launchpads.

Notice those are the same categories Robinhood's income statement is pointing at. Perps and event contracts are exactly where the money moved. Two completely different vantage points, one from a trader and one from a public company's revenue mix, arriving at the same place.

All of the above is analysis, which has to be turned into positioning, and it's worth running through because it maps conviction level to instrument rather than just saying "buy the trend."

If you think memecoins keep running but you have no edge picking winners, buy the infrastructure that earns on the volume regardless of which coin wins: UNI, RAY, JUP. If the mania stays concentrated on Robinhood Chain, ARB gets a 10% share of that as the Orbit chain underneath it. If it's Solana memes dominating, that's SOL.

If you believe in the trend but want more exposure than infra gives you, the launchpads are the next rung out, since every new token pays them first: PONS, STONK, RAY's LaunchLab, and PUMP. More correlated to the meme cycle than the L1 tokens above, but still one step removed from picking individual coins.

If you're comfortable being fully degen, that's the point where you're picking the specific meme winning on whichever chain has attention that week. Different risk category entirely, sized accordingly.

The LP angle is the part that doesn't get talked about enough. Tokenized stock/USD pairs are, in his framing, an unusually clean yield: arb bots keep the onchain price pinned to Nasdaq, so impermanent loss stays small while fees stay high, and every meme-to-stock rotation this section has been describing swaps through that same pair. Long-tail yield without the long-tail drawdown, and Beefy and other aggregators are already building vaults to make it closer to set-and-forget.

None of this is a recommendation; it's a way to sort the trade by how much conviction you actually have. The failure mode is buying the degen tier with an infra-tier thesis.

None of the above survives contact with a bad inflation print, and this is the last week of inflation data before the September Fed meeting.

Friday is the day. A hot CPI and every rotation argument above gets repriced in an afternoon, because it all runs on the same liquidity assumption. A soft one and phase two gets its funding.

Does the Robinhood mix hold? One quarter is a crossover, three quarters is a business model. The next print tells you which.

Does anything follow $MEME and $BONER? Two CEOs is an anecdote. Ten is a channel, and the tokenized equity trade stops being theoretical.

Position size into Friday, not after it. The rotation Kyle describes may well be right and still be unrecoverable if you're sized for a benign CPI and don't get one.

One honest caveat on all of it: phase-two calls are much easier to make than to time. "Rotation into fundamentals" has been the correct-sounding answer in every cycle, including the ones where it took another eight months to arrive.

The last week of inflation data before the September Fed meeting.

Apart from that, we have a couple of events worth keeping an eye on.

The read: CPI and FOMC aren't just the two biggest lines, they're a four-day window that swallows everything landing inside or just after it. AERO's deployment falls right in the middle of it. XPL's unlock lands the week after, when the market is still digesting the Fed and liquidity is thinner than normal — which is exactly when a record-sized cliff does the most damage.

If you're carrying size into Thursday, that's a decision you make today, not at 8:29 ET that morning. The people who get hurt on CPI days are rarely wrong about direction. They're wrong about how much they were holding when the number landed.

One of those once-every-six-month trades. He won't buy it now, this is speculation parabola, uPNL is high, and a flush is coming at some point even if the cash flows keep running for a while.

The value here isn't the ticker, it's the discipline: when you miss a big move, the next job isn't chasing it, it's getting aggressive on the flush when it comes.

Started adding on the ~30% pullback to $0.004. His case: price-to-revenue is still bluechip-tier for DeFi, buybacks annualize to 10% of FDV, and unlocks are manageable at roughly 15% of supply a year. He thinks onchain activity is rotating back to Solana after Robinhood Chain pulled attention away, and he's watching for Pump.fun to announce a launch-pairs expansion onto Robinhood Chain itself as the next catalyst.

Sizing is precise, not a single entry: building at $0.004, DCA down to the low $0.003s if it comes, targeting a "LIT fractal" back above $0.008 for 100%+ upside.

Added again on the first H4 trend retest plus an RSI reset, after an H4 bear divergence played out exactly as flagged. Invalidation is acceptance under the H4 trend; if that breaks, H12 support at $0.15 comes into focus.

Context worth knowing: this follows what Paradise called the biggest ETH runner of his account last week, already 90% closed for profit. This is a second position built on a fresh technical trigger, not a hold-and-hope.

Back at the desk after a night off, with ETH as the main event: still liking the compression into $2,520, and calling 2.8-3K "inevitable" once that level gives way on acceptance. Everything else on his list is framed as a follower trade, MORPHO trading above trend and likely to go if ETH goes, ENS holding above the support he'd flagged, LDO and JUP starting to break out.

The structure is the takeaway: one high-conviction trigger, and three names sized around whether it fires.

Already up over 106% on a 25x perp. Taking partial profit at $93, the upper quarter-range extension from the prior range. Main target is just above the $100 round number, the full range extension, and he's leaving a runner if price gets there.

Worth noting for anyone who wasn't in this one: he's de-risking into strength on a name already at all-time highs, not adding.

Free Skool Community

For more behind-the-scenes, including AI prompts, workflows, and research, join Miles’ free Skool community here: