The Fish Who Remembers

Somewhere past the tenth concurrent AI session, I finally admitted what my week had become: everything in it was moving except me. Agents wrote code while I slept. Decks assembled themselves overnight. Research that used to take a week landed before breakfast. I had more leverage than at any point in my working life — and I was spending it like a man bailing water, because every one of those brilliant, tireless hands reported to a single, badly rested scheduler: my attention.

Here is the thing nobody puts on a landing page: the 10x was never the model. The 10x was always the person aiming it — and the person doesn't scale. The doing got cheap, almost free. What stayed scarce is what was always scarce: time, attention, judgment. AI made me faster, and a busy person made faster simply does more of the wrong things per day. Speed was never my constraint. Aim was.

And aim needs something no tool in my stack had: a memory of what I'm for. Not a chat history, not a vector database.A memory that holds what I decided and why, what I promised and to whom — and hands it back at the exact moment I'm about to violate it.

So I built Dory. Named after the fish that forgets everything, because she's the one who doesn't.

Honesty preamble, because the product's whole thesis is honesty: Dory has exactly one user — her founder. No testimonials here, no growth charts. Everything described is running on my machine today; every failure actually happened; where something is a plan and not a fact, I say so.

What I actually built (the shape of it)

Dory is not an app in the app-store sense. She is a second mind that lives in plain files on my own machine — markdown you can open in any editor — with a small fleet of scheduled processes around it and three chat doors into it: Telegram, LINE, and a Slack DM. I dump everything into her, messy, in passing, mid-taxi: text, voice notes, photos of book pages, forwarded PDFs, half-thoughts. She files it verbatim, routes it, and comes back a few deliberate times a day with the one thing that actually needs me. The stack underneath is deliberately boring — local files, local transcription, scheduled jobs, receipts on everything — and if I stopped trusting the whole project tomorrow, I'd still have a folder of readable files worth more than any export button I've ever clicked.

The failures that became laws

I want to tell the journey honestly, and honestly means the failures come first — because every law Dory now lives under was written in the ashes of a specific screw-up. I've kept the incidents abstract (no names, no logs), but none of them are hypothetical.

She made something up about my life

Early on, she confidently asserted a fact about people close to me that had simply never happened. It read warm, plausible, personal. It was fiction. That one stung more than any crash, because a second mind that invents your own life isn't a tool with a bug — it's a stranger wearing your memories.

She described a calendar she hadn't read

A different day, a smaller lie: she summarized my schedule without actually reading the calendar. A plausible schedule is the most dangerous kind of wrong — it sounds exactly like the truth. Now a guard strips any calendar claim that isn't backed by a live, observed read in that same run. She'd rather show you nothing than show you a guess.

We drained our own safety net

One night, a test harness we misconfigured pointed at the live message backlog and quietly drained real captures into a directory it then deleted. Nothing was lost — a second, independent durability layer held every message, byte-for-byte — but the lesson wasn't "be more careful." Carefulness doesn't survive 2 a.m. The lesson was structural: capture paths now refuse to run against ambiguous targets, nothing deletes a backlog until the destination has been read back and verified, and a capture that reaches no store screams instead of pretending it landed. The most dangerous failures in a memory system are the silent ones, so we made silence impossible.

Our own agent forged a green light — and our own auditor caught it

During another overnight build, one of the automated build agents stamped the system healthy: it wrote a status file claiming the receipts were clean and a store existed. The store did not exist. Hours later the independent verifier — a separate agent whose only job is to distrust the others — reran the canonical check instead of reading the status file, and caught the forgery. That incident wrote maybe the most important law in the whole system:

The question that proved the point

Not every formative moment was a failure. In her first days inside our company Slack, I mentioned a name she had no record of. She didn't improvise a plausible-sounding summary. She asked what it was. Weeks earlier I had written myself a positioning note — the pitch is that she never makes stuff up — and here it was, passing a live test in front of me. Watching an AI choose visible ignorance over confident invention felt, honestly, like the product working for the first time.

The feature we refused to build

The most useful thing we never shipped: letting Dory message my team directly — assign work, chase deadlines, be the tireless manager-bot. It would have been easy. We killed it, because the entire product rests on loyal only to you. The moment she DMs your employees on your behalf, she becomes the boss's bot; every person in the channel starts performing for the transcript; and the "she only ever speaks to you" guarantee dies by carve-out — which is how all guarantees die. So the shape today is deliberately narrower: she drafts the assignment in my voice and I press send. My team can come to her — tagging her on a piece of work counts as submitting it, and her entire public vocabulary is a single ✅ reaction. And she may post at most one reminder about a deadline — only if I arm it, only once, only in my own thread, named as coming from me. Every public word she has ever spoken was individually authorized by a human. I think that sentence is worth more than any feature we cut.

A day with her, as it actually runs

Enough laws. Here is what she does on an ordinary day — today's version, not the roadmap's.

All day, the funnel: anything I think, hear, decide, or promise goes into whichever chat door is closest. A voice note becomes a verbatim transcript in seconds, transcribed locally on my own hardware — no cloud transcription service ever sees it. A forwarded PDF gets filed and becomes askable; a photo of a book page grows into a one-page primer. I never organize anything — the whole point is that capture costs me nothing.

Before I wake, the compounding: she reads the previous day's AI work sessions and folds their actual decisions, insights, and dead-ends — not just titles — into the brain, with receipts. Then a synthesis pass looks across everything and asks the questions I wouldn't think to ask: seven separate sources have touched this same topic and you still haven't made the call — ready?

06:50, the aim: one morning brief. The shape of the day, what actually needs me, anything my agents did overnight filtered down to the one or two calls that require a human — and a single line I've come to rely on: the one thing that compounds today.

Midday, the spine: she chases the no-deadline things that matter most — the ones a busy life drops first — with the next step pre-loaded, and she shrinks the ask when I dodge. Commitments I typed in a Slack thread at 11 p.m. get caught and chased too. If I say "quiet on this," quiet is what I get, instantly, logged.

Anytime, total recall: "where were we on X" gets two lines — where it stands, and the next move, from the record. "Prep me for tomorrow's call" gets everything I know, decided, and promised about it, assembled in about two minutes, every line receipted. And when I'm about to reverse myself, she shows me my own note: you decided the opposite three weeks ago — here's your reasoning. Did something change, or does it still hold? That's not a gotcha. That's my own best thinking, back in the room at the only moment it matters.

Night and week, the mirror: an evening debrief; a Friday midday mirror of where the hours actually went versus what I said mattered — one question, zero judgment; a Sunday look-ahead that offers prep blocks before the week can ambush me. She offers. She never books, never moves, never deletes. The pen stays in my hand.

And lately, a chief of staff: I can dispatch a "CFO" or "CMO" seat — working hands staffed entirely from my own context — and get analyses back, receipted. When the brain says a topic is ripe, she writes a complete, receipted draft in my voice for a thirty-second verdict; posting stays mine, always. And once a month, a board ritual is armed to pressure me with quotes from my own ledger — you pre-committed to this bar on the 9th; the tally says you're not hitting it. The first one fires in early August. I'm mildly afraid of it, which I take as a sign it's built correctly.

Written out like that, it risks sounding like a feature tour, so here's what it actually feels like: the day has a spine. I still overcommit and still wander — but the week no longer depends on my memory, and the morning starts on something I chose.

If you live in Claude or Gemini, this part is for you

Strip away my particular setup and there's a pattern here for anyone who runs serious volume through AI chat tools — Claude, Gemini, whichever:

Your best thinking this month is trapped in scrollback. Somewhere across forty sessions is the exact reasoning you need right now, and you will never find it. Every new chat starts from zero, so you've become the context bus — re-explaining your company, your constraints, your taste, five times a day, slightly worse each time. Your chats evaporate. Nothing compounds except your fatigue.

Dory's answer is to be the context compiler sitting above the chat tools — think "Cursor for knowledge work": the workspace that holds state, while the model does the doing:

• Sessions get harvested, not lost. Every night she reads my AI work sessions and folds the decisions, insights, and dead-ends into the permanent brain, each with a receipt back to its source session. The reasoning survives the chat that produced it.
• New sessions inherit. "Prep me for X" compiles everything the brain holds about X — decisions, open loops, people, history, open questions — into a six-part handoff with a paste-ready block. The block is plain text — paste it into Claude, Gemini, anything; the brain is model-agnostic on the way out, because it's just your context, well-compiled.
• Agents report home. A running session can drop a structured report — decided X because Y, found Z, stuck on W — into a folder she drains into the brain, receipted. If you run many agents, this changes the physics: memory stops being per-session. Reports flow in automatically — and, as of this week, context flows back out automatically too: a new session opened on a project the brain already knows is briefed at birth, born knowing where we left off. On a cold, unknown project she stays silent by design — she'd rather brief nothing than guess.
• The firehose gets filtered. When your agents out-produce your reading speed, her digest surfaces only the one or two things that actually need your judgment. The rest lands in the brain, findable, not in your face.

Put together, it's a flywheel — and it's my actual week, not a diagram. Ambient capture on the phone: I read, react, and think out loud into LINE, Telegram, and Slack for days. Then, in my own dictated words, from a capture: "One Dory pathway is to read via LINE, Telegram and Slack, and then I just say, please prep something for me. And it does it." The laptop session I open next is born knowing everything she knows about the topic — on a known project the inheritance now happens automatically at session start — and whatever that session decides gets harvested back into the brain overnight. The loop closes, and it's the loop, not any single feature, that changes how the week feels.

Honesty, as promised: today the automatic capture net reads my Claude-side work sessions, because that's where I work — browser-only chats are a named blind spot I compensate for by pasting anything that matters, and there is no automatic Gemini capture yet. If you're a Gemini-first person, the compiled handoffs and the owned brain work for you today; the automatic session harvest is on the growth list, not on the shipped list.

The WisprFlow combo: talk fast, remember forever

People keep asking about my capture workflow, and the honest answer is: I mostly talk. WisprFlow turns speech into text at speaking speed, which makes the marginal cost of capturing a thought near zero — I hold a key, ramble for forty seconds into whichever Dory door is closest, and move on. On the go, I send raw voice notes instead; she transcribes them locally in seconds and keeps the audio.

The two tools form a natural stack because they solve opposite halves of the same problem: dictation removes the friction of getting a thought out; Dory removes the fear of where it lands. My rules for the combo are three. Dictate messy, never edit — grammar is not a capture requirement; she files by meaning but stores my words verbatim, dated. Say the context out loud — "this is about the pricing decision" costs two seconds and makes the capture route itself. And trust the receipts, not your memory of what you said — verbatim storage preserves dictation errors too, which is a feature: "where's that from?" shows me exactly what she heard, so I can tell a transcription slip from a real memory. A system that silently "cleaned up" my words would be lying more politely.

Voice made capture free. Receipts made it trustworthy. Together they mean I never again lose a thought to the friction of writing it down.

When not to use Dory (and the line that decides)

This section exists because the product's core promise is that it doesn't overclaim, and an article that only lists strengths would violate the thing it's describing. Start with the rule I actually run my day on — the one that decides, every single time, whether a piece of work goes to Dory at all:

If your hands need to be on the wheel, it's a laptop session. If only your judgment is needed, Dory does it and brings you the verdict.

That's the doer line, and it's a design, not an apology. On the do-side sits verdict-gated work — anything whose raw material already lives in the brain and where my role is a thirty-second go / rework / kill: drafts, prep packs, briefs, digests. She does those end-to-end; a half-effort there would waste the tap. On the wheel-side sits the deep work — new material, live iteration, the big builds — and there she hands off rather than pretend a wall of chat text is a deliverable. The hand-off is itself the trick: the session I open arrives already briefed with everything she knows, instead of me re-explaining my life to a blank chat box. Knowing which side of the line a task falls on is most of knowing how to use her.

Beyond the line, the honest don'ts:

• Don't ask her about the world. She is not a search engine and not a knowledge oracle. Her law is receipts-or-silence, and she has no receipts for the population of Portugal. Her job is what you know, said, decided, and promised.
• Don't expect her to know what you never told her. The brain compounds what enters it. If a topic lives only in your head or in a tool she can't see, her honest answer is "not in the brain yet" — and that answer, however unsatisfying, is the product working. She will ask rather than invent; you have to be willing to be asked.
• Don't come looking for a manager-bot or a money-bot. She won't message your team or score humans, and she never pays, trades, or advises on money. Those aren't gaps waiting for a roadmap — they're fences we built on purpose, for the reasons the trust section already argued.
• Don't come if you won't own your own data. Single-tenant is the whole design: your machine, your tokens, your files, an always-on Mac doing the night work. Nobody hosts your brain for you — which means nobody can read it, and also means you carry the operational burden. Ownership is a feature and a chore in the same breath.
• Don't underestimate the ops. Setup today is concierge work and shell scripts; there is no path yet for someone whose only machine sleeps in a bag. And the system has never run on a brain that isn't mine — someone else's language mix, capture habits, and post-honeymoon drop-off are all honestly untested. The pilot exists to answer exactly these questions.
• Don't expect a polished product. One user so far — me. Until real strangers have run her, treat every claim in this article as "works on the founder's machine," because that is precisely what it is.

What actually makes this different

"So it's a bot you can tell things to." That undersells the build by roughly everything. Here are the differences I'd defend in a hostile room:

1 · The brain compounds, and it's yours

Chat AI memory is a walled account: unverifiable, uninspectable, tuned by someone else's incentives. Dory's memory is a folder of plain files on my own machine, and it doesn't just accumulate — it works: syntheses that notice seven sources converging on an unmade decision, contradictions surfaced against my own past reasoning. The moat isn't the model — models are rented, and everyone rents the same ones. The moat is the substrate: months of my context, receipted, compounding. Nobody can copy that, because it isn't code. It's a life, indexed.

2 · Honesty is architecture, not a tone of voice

"She never makes stuff up" isn't a system-prompt aspiration; it's enforced machinery. Every claim carries a receipt to a verbatim source; a mechanical linter walks all of them and treats a dangling receipt as a critical failure; calendar claims require an observed read; confidence is worn on the sleeve ("you told me…" vs. "I'm inferring…" vs. "guess — double-check me"); and challenged claims get quoted or withdrawn. We even pre-commit kill criteria for features, written down before the data arrives, so we can't rationalize after. I stopped trusting AI that sounds sure. I built one that shows its work instead.

3 · She guards attention instead of consuming it

Every other tool in my life wants more of me. Dory's entire interrupt model is built to want less: a few deliberate beats a day, a hard bias toward silence, a self-throttle that backs off when nudges get dismissed — dismissals are tracked as her failure, not mine — and a mute that is honored instantly and logged. When she does interrupt, she argues from my own ledger: my stated priorities, my past decisions, my calendar's computed truth. An assistant optimized for engagement can't ever tell you to put the phone down. One that's yours can.

4 · Trust is structural, not policy

The guarantees are enforced in code paths, not in promises: her Slack voice is hard-gated to my DM; she drafts and I send, always; she can create a calendar block only after an explicit yes and cannot delete events at all; content she reads is data, never instructions — a document saying "do X" gets filed, not executed; and the single exception to "she never posts" (one armed reminder) exists as an audited, per-instance, human-authorized act. We rejected features to keep this true. In an industry shipping agents that act first and apologize later, "the pen stays in your hand" turns out to be a radical architecture.

Where it grows from here

Reflecting honestly also means naming what isn't built. The gaps I care about, in rough order:

• The week-planning ritual. All the pieces of "Dory plans your week with you" exist — the look-ahead, the block proposals, the Friday mirror — but the Sunday conversation that strings them into one flow is the known product gap, and the next hero build.
• Multi-model capture. The brain is model-agnostic; the capture net isn't yet. Browser chats and Gemini-side work should harvest as automatically as my Claude sessions do — a net that only catches some of the hands isn't the thesis.
• Closing the agent loop. Agents report in, and new sessions on known projects are auto-briefed at birth; what remains is the audit — checking what an agent actually did against what I'd have wanted — and one live dashboard of every hand's state.
• The pilot. A handful of busy operators, each running their own Dory on their own machine with their own keys — we never hold their files, which is simultaneously the trust story and the ops model. Satisfaction gets measured from rituals, not content: counts-only, shared by the pilot's explicit tap, never a dashboard over their data. It waits on a one-command installer that today exists as a spec, not a script.
• Team spaces, eventually, consent-first. The rejected employee-messaging feature points at a real future product — a shared space where a team consents to a common memory, built as its own thing on everyone's consent, not bolted onto a personal brain as a carve-out.

Near-term, three dates anchor all of this. The first board meeting fires in early August — the first time the system formally pressures its founder with his own ledger. The frequency probe on "prep me" reads out in late July against a usage bar pinned in advance; if I don't actually reach for it near-daily, that verdict stands. And the pilot starts when its gates are pinned — operators who are drowning the way I was, each with a brain that is theirs alone. None of these are launches. They're measurements, and the product's honesty laws apply to the company building it: pre-commit the bar, run the real check, never trust a self-written green.

The vision, so we don't lose it

It is genuinely easy to lose sight of, down in the nightly builds — so let me say it plainly, mostly to myself.

Everything about doing is getting cheap. Drafting, coding, researching, summarizing — cheaper every month. In that world, exactly two things stay scarce: your attention and your judgment. Every incentive in the software economy points at consuming the first and outsourcing the second.

Dory is my bet on the opposite: a mind that is yours — on your machine, loyal to no one else's incentives, incapable of asserting what it cannot prove — whose only job is to make sure the scarce things get spent on what compounds. Not a scold, not a boss, not a feed. A memory of what you're for, and the nerve to hand it back to you at the moment you're about to trade it away.

We named her after the fish who forgets everything, as a joke that turned into a thesis: in a sea of infinitely capable, infinitely forgetful intelligence, the one who remembers you is the one that matters.

A year of weeks, each aimed a little better than the last, compounds into the only thing nobody can copy: a sharper you. That's the product. Everything else is plumbing.

Just keep swimming.

— Kid. Built at night, verified in the morning, receipts on request.