Most Agents Forget Everything by Morning. The One That Keeps Score Will Quietly Lap Them All.
June 30, 2026 · Ray Kurzweil, The Singularity Is Near~6 min read
Most AI agents wake up with amnesia. Run the same one on the same job a hundred times and on the hundred-and-first it will still botch the opening move exactly the way it did on day one — the session closes, the context window flushes, and everything the thing worked out drains away with it. We stopped noticing, because that is just how agents are. Then in June a small lab shipped one that doesn’t do that, and the part worth your attention isn’t that it got smarter. It’s that it stopped forgetting.
The agent with anterograde amnesia
Picture what “the session ended” actually costs. An agent spends twenty steps learning the quirks of your codebase — which test fails at random, which API lies about its rate limit, the one incantation that makes the deploy script behave. It finishes the job. Then the window clears and all twenty steps are gone. Next week the same task lands and the agent re-derives the whole thing from scratch, tripping over the flaky test again, getting bitten by the same rate limit again. It is Groundhog Day with a token budget, where every run is the first run. Smart, sure — but smart the way a brilliant new hire is smart on their first morning, every single morning, until the heat death of the universe.
What Hermes writes down
Nous Research’s Hermes Agent breaks the loop with a move so plain it’s almost a letdown: it writes the win down. When a task finishes with roughly five or more tool calls, a quiet background process summarizes the trajectory — the actual sequence of moves that worked — into a small Markdown file and drops it in a folder on disk. Next time, that file is just sitting there: a reusable skill the agent reaches for instead of solving from zero. The current build ships with 118 of them, and they’re plain text, so you can read one, edit one, delete a bad one, commit them to git like any other file. It runs on an entry-level VPS and answers in your Discord. None of that is the headline. The headline is that the lesson got saved somewhere the next session can find it.
Step back and the timing gives the game away. June 2026 was a flood of open models — Zyphra trained an eight-billion-parameter model from scratch on AMD silicon instead of Nvidia’s and put it out under Apache-2.0; Nvidia shipped Cosmos 3; Anthropic released an Agent SDK. Skim the month and the story reads “models everywhere, cheaper, more open, a little smarter.” But raw smarts-per-query was never the bottleneck for most real work. The bottleneck is that the system forgets between this Tuesday and next Tuesday. Hermes is a bet on the boring half of that problem, and the boring half is where the leverage hides.
Most AI agents flush their context when a session ends, so every run starts from zero. Nous Research’s Hermes Agent instead compiles a successful task trajectory (roughly 5+ tool calls) into a permanent Markdown skill saved on disk — 118 ship with the current build — so the next run reuses it and adds more, an ability curve that compounds rather than resets. Framework: Ray Kurzweil, The Singularity Is Near — the law of accelerating returns, where a system that recycles its own gains rides a compound curve. This is narrow skill accumulation, not recursive superintelligence. Figures as reported; popular-science interpretation.
Why compounding beats horsepower
Ray Kurzweil built The Singularity Is Near on one engine he called the law of accelerating returns: when a system can use its own output to build better tools and fold those gains into the next round, it stops moving in a straight line and starts riding a curve. The distinction is the entire game, and it lands in your gut the second you frame it as money. A model that’s 10% smarter is a one-time raise — nice, spent, gone by the next benchmark. A system that keeps 10% of what every task taught it is an interest rate. Strength is a number; interest is a slope. Over any horizon worth caring about the slope wins, and it isn’t close. Kurzweil’s other warning rhymes here: we read these curves with linear eyes, so we overrate what compounding does this quarter and badly underrate what it does in three years.
That is the real reason a remembering agent matters more than a marginally cleverer one. The clever one hands you a better answer today. The remembering one hands you a worse answer today and a better one every week after, because every solved problem leaves a deposit. Call it the memory interest rate. The frontier labs are sprinting to raise the first number — bigger model, higher score. Hermes is quietly compounding the second.
The ceiling, honestly
Here is where the hype wants to bolt ahead, so plant your feet. What Hermes does is not recursive superintelligence. It is not rewriting its own weights or getting fundamentally better at reasoning; it caches procedures — narrow, bounded, sometimes wrong skills that a human still has to prune. And compounding cuts both ways: cache a mistake and you have automated the mistake, faithfully, every run. The phrase “self-improving” is doing a lot of quiet work in the marketing, and it must not be allowed to smuggle in “singularity.” Kurzweil’s curve is a claim about the long arc of many systems over decades, not a promise that this June’s Discord bot is rung one on the ladder to godhood. The honest version is smaller and far more useful: a system that compounds beats a system that resets, even when both are nowhere near superhuman. That isn’t the rapture. It’s just better engineering — and better engineering is the part that actually ships.
What this means for you
If you’re deciding what to build on or bet on, quit shopping for the cleverest model. That edge has the shelf life of milk; next quarter’s release sours it. The durable edge is the system that keeps what it learns — the one whose hundredth run starts where the ninety-ninth left off instead of back at zero. Then turn the same lens on yourself, because you are also an agent that mostly resets. You solve something hard, ship it, and six months later re-derive it badly, because the trajectory lived only in a context window that flushed. The fix is unglamorous and it’s identical to the agent’s: write the move down, not just the result. Keep the skill, not the trophy. You won’t feel the difference next Tuesday. You’ll feel it the Tuesday a year out, when a problem that used to cost you a day costs you ten minutes — because past-you left a file where present-you could find it. Don’t start from zero. Almost nobody does the boring thing, which is the entire reason it compounds.
A model that’s 10% smarter is a one-time raise. A system that keeps 10% of every lesson is an interest rate — and over any horizon that matters, the slope beats the number.
Bet on the system that remembers, then go be one yourself: keep the move, not just the result.
Source: Ray Kurzweil, The Singularity Is Near (the law of accelerating returns — a system that recycles its own gains into better tools rides a compounding curve, not a linear one, and we habitually underestimate the long run). News pegs: Nous Research’s Hermes Agent, which compiles successful task trajectories of roughly five or more tool calls into permanent, on-disk Markdown skills (118 bundled in the current build) and runs on an entry-level VPS with Discord/Slack access; and the June 2026 open-model wave (Zyphra’s Apache-2.0 ZAYA1 trained from scratch on AMD, Nvidia’s Cosmos 3, Anthropic’s Claude Agent SDK), per devFlokers’ roundup. This is narrow skill accumulation, not recursive superintelligence. Popular-science interpretation; figures are as reported.
技术
大多数 AI agent 睡一觉就忘光——那个把所学攒下来的,正悄悄把它们甩开一圈。
2026 年 6 月 30 日 · 雷·库兹韦尔《奇点临近》约 6 分钟
大多数 AI agent 都是「睡一觉就失忆」。同一个活儿让它干一百遍,干到第一百零一遍,开头那步它照样能搞砸,跟头一天一模一样——会话一关,上下文窗口一冲,它刚摸出来的门道全跟着水一块儿流走。我们见怪不怪,因为 agent 本来就这德行。可六月里,有家小作坊放出一个不这样的,值得你抬头看一眼的,不是它更聪明,是它不再健忘。
たいていの AI エージェントは、寝て起きると記憶を失っている。同じ仕事を百回やらせても、百一回目の出だしを、初日とそっくり同じように仕損じる——セッションが閉じ、コンテキストの窓が流され、せっかく掴んだ勘どころが、まるごと一緒に流れ去る。私たちが気にも留めないのは、エージェントとはそういうものだからだ。ところが六月、ある小さな工房が、そうではない一台を世に出した。目を上げる値打ちがあるのは、それが賢くなったことではない。忘れなくなったことだ。
「覚えた端から忘れる」職人
「セッションが終わった」が、実際どれだけを焼き捨てるか考えてみてほしい。エージェントは二十手かけて、あなたのコードベースの癖を掴む——どのテストが気まぐれに落ちるか、どの API が口先の流量制限と実際とで食い違うか、デプロイスクリプトを大人しくさせる、あの一言の呪文か。仕事は片づく。窓が消え、その二十手が消える。来週また同じ仕事が来れば、一から推し直し、あの気まぐれなテストをまた踏み、同じ流量制限にまた噛まれる。トークン予算つきの『恋はデジャ・ブ』だ。どの回も初回。賢いのは本当だ——だがそれは、飛び抜けて優秀な新人が、毎朝、出社初日のままでいる賢さだ。
Hermes は、その一手を帳面に書きつける
Nous Research の Hermes Agent がこの輪を断つやり方は、拍子抜けするほど素朴だ。勝った一手を、書き留める。タスクがおよそ五手以上のツール呼び出しで片づくと、裏で静かな処理がその軌跡——実際に効いた手順の連なり——を小さな Markdown ファイルにまとめ、ディスク上のフォルダにそっと置く。次は、そのファイルがただそこにある。再利用できるスキルとして、エージェントはゼロから解き直す代わりに、それへ手を伸ばす。今の版には 118 個が同梱され、どれも素のテキストだ——読めるし、直せるし、できの悪いのは消せるし、ほかのファイルと同じように git に積める。入門級の小さな VPS で動き、あなたの Discord で返事をする。どれも見出しではない。見出しは、その一課が、次のセッションが見つけられる場所に保存された、ということだ。