You had a conversation last Tuesday that mattered.
Maybe it was the indemnity cap on the Acme term sheet. Maybe it was the way your head of engineering paused a beat too long when you asked about the rewrite. Maybe it was a throwaway line from a podcast that's been rattling around in your head ever since.
A week later, you want to pull on that thread. You open your AI assistant. You ask.
It has no fucking idea what you're talking about.
It has never heard of Acme. It doesn't know you have a head of engineering. It cannot tell you what you thought was interesting last Tuesday because it was not there last Tuesday. It does not persist between sessions. Even if it did, it would never have been allowed near your actual data. And even if that, the round-trip to wherever it lives takes long enough that you'd have time to make a coffee, drink it, and start writing the answer yourself.
So you paste. You paste context. You paste more. You wait seven seconds while a spinner does an impression of thinking. The answer comes back plausible, generic, and subtly wrong in a way you can't be bothered to correct. You close the tab. You open Slack. You ask a human.
This is the state of the art in late 2026. An entire industry looked at this scene, took a long breath, and concluded that the answer was a more articulate chatbot.
We don't.
Five failures. Pick your favourite.
Let's be honest: the status quo sucks.
The promise was a personal AI that knew you. What shipped was five compounding failures stacked into a trench coat and sold as a subscription.
01Amnesia
Every chat starts from zero. The model is a brilliant stranger you keep re-introducing yourself to. "Context windows" help within a single conversation and then evaporate the moment it ends. What you told it last week is gone. What it told you last week is gone. What you committed to, what you decided, what you noticed, what you were wrong about — all of it, gone.
A human assistant with amnesia this aggressive gets fired on Tuesday afternoon. Which — you may have noticed — is also the day your AI forgot about Acme.
The chatbot vendors' answer to this is a "memory" feature that retrieves a handful of summarised facts from a previous session and stitches them into the prompt. This is the software equivalent of giving the stranger a Post-it with your name on it. It is not memory. It is a prop. The stranger is still a stranger. You're just saying hi louder.
02The Librarian Tax
The existing answer to amnesia is a second brain — Notion, Obsidian, Roam, Logseq, Tana, Heptabase, pick your poison. Each one politely asks you to become an unpaid librarian of your own life. Tag the thing. Link the thing. File the thing. Maintain the ontology. Don't forget the ontology. Forget the ontology and it's worse than if you never started.
The tools are beautiful. Some of the best software of the last decade, honestly. They also fail for the 99% of people who cannot, on top of running a company and a household and a calendar, also run a personal Dewey Decimal System. Look at the people for whom second brains actually work — really work, not in the YouTube-tutorial sense — and you will find that they have turned knowledge management itself into a hobby. They have spreadsheets about their notes. They have notes about their spreadsheets. They have strong opinions about bi-directional linking. Second-brain people are the vegans of knowledge work — you find out within five minutes of meeting one. You did not ask.
Good for them. If you're reading this, you're probably not one of them. Also good.
The Librarian Tax is the price of admission to every current tool. Skip the filing, skip the tagging, skip the curation — and your second brain becomes a worse Notes.app. Do the filing, tag the tagging, curate the curation — and you have a new part-time job. No pay. No benefits. You wanted this.
03The firehose
The next-generation answer to the Librarian Tax is to automate the ingestion: just plug every integration into your second brain and let the AI sort it out. This is worse.
Your iMessage history contains your therapy. Your email contains your board drama. Your calendar contains the meeting nobody was supposed to know about. Your browser history contains whatever you searched at 2am last Tuesday while you couldn't sleep. Sucking all of that, indiscriminately, into a vector database and letting the model retrieve from the whole pile is not a feature. That's NSA-ing yourself. Bold move.
The tools that do this — the Rewinds, the Mems, the we-index-everything crowd — treat ingestion as a plumbing problem. Pipe the data in, model sorts it out, ship the thing, raise the round. Except the model does not know which of your conversations is protected by a non-disclosure agreement. It does not know which messages contain the private life of someone you love. It does not know the difference between a draft you abandoned and a document you shipped. It cannot make those judgments, because those judgments require knowing things the model structurally cannot know.
And even if you didn't care about any of that — there's a deeper failure. These tools don't know that David in iMessage is David Smith in your email is your head of engineering. They see three different strangers who happen to share a first name. Resolving entities across sources, merging identities, cross-verifying claims against each other — that's a recursive problem, not a plumbing one. It is the hard part of the whole enterprise. Nobody ships it. You end up with a disconnected mess of strangers in a trench coat.
What's missing is an intermediary — a layer between the raw firehose and your knowledge substrate that lets you curate, scope, and veto what gets ingested. A filter with your values in it. Underneath it, a graph that actually resolves who's who.
The fact that nobody ships this should tell you something about who the current tools were built for. Not you.
04This is your life. Act like it.
To get useful output from the cloud tools, you upload your most sensitive context to someone else's servers. Their terms say they won't train on it, probably, unless they change the terms — which they will, quietly, on a Friday afternoon in July.
Every serious person has felt the pause before typing a name, a number, a real thing. You weigh whether typing it is worth the help. You type it anyway, feel vaguely complicit, move on. This is fine. Everything is fine.
The trick of the last ten years of SaaS was convincing a generation of professionals that their data was not really their data. It was user-generated content. It was input. It was telemetry. It was fuel, training data, product. Your relationships, your notes, your thinking — recast as raw material for somebody else's margin. Quarterly earnings calls referenced you as a retention cohort. People noticed. They typed it in anyway. The tools were useful, and the alternative was no tool at all.
This is your life. Act like it.
There is now an alternative.
05Speed — or rather, the lack of it
This is the one nobody talks about. Modern AI assistants are so slow.
Seven seconds to answer "what did we decide about pricing." Thirty seconds if you want an answer worth reading. Three seconds to open the app. A full second of network latency on every keystroke because the compute is in Virginia and you are not in Virginia.
Cloud chatbots cannot be fast. There is a model running somewhere else, a queue in front of it, a load balancer in front of the queue, and a transoceanic cable in front of the load balancer. This is not a bug they'll fix in the next version. It is the physics of the architecture everyone picked.
The industry has quietly redefined acceptable latency. What was unacceptable in 2012 — three seconds to open an app — is somehow fine in 2026, because AI. No. Slow tools do not become fast tools by adding features. They become slower tools with more features. And slow tools shape behaviour: you use them less, you trust them less, you route around them.
A tool you wait for is not an extension of your working memory. It is an errand.
The failures are not independent. They feed each other.
Amnesia forces you to paste context. Pasting is tedious, so you organise the context — Librarian Tax. The Librarian Tax is too expensive to pay, so you reach for automated ingestion — firehose. The firehose is too sensitive to run on your own hardware, so you hand it to the cloud. The cloud is slow and surrenders your privacy. You give up. Open a text file. Promise yourself you'll fix it next quarter.
Five failures. One loop. Everyone running it together.
The diagnosis
The industry spent the last three years convinced that the product was the model. Scale it up. Train it harder. Benchmark it against the competition. Pay a million dollars a month to whoever's ahead this quarter. Ship the chat window. Charge for the chat window. Churn the customers who figured out you didn't mean it. Raise again.
The model was never the product. Models are commodities. They get cheaper, faster, and more interchangeable every month. Anyone who told you otherwise was trying to sell you a subscription.
The product — the thing that is actually scarce, the thing that actually compounds — is your context. Who you know. What you decided. What you noticed. What you committed to. What you learned. What you built. That context is not in the model. It cannot be in the model. It lives across your email, your calendar, your meeting transcripts, your daily notes, your messages, your browser history, your documents, your head.
The failures above are all symptoms of the same category mistake: treating context as something that happens inside a session, inside an app, inside somebody else's data centre. It is not. Context is the substrate. Everything else — the model, the chat window, the integrations, the subscription — is a thin coat of paint on top of it.
You cannot fix this loop from inside it. You cannot add memory to a chatbot that doesn't run on your hardware. You cannot eliminate the Librarian Tax by piping more data into a system with no filter. You cannot protect your privacy by asking the vendor nicely. You cannot make a cloud service fast without putting the cloud on your desk.
Here's what the status quo costs you. A chatbot without your context is a machine for mean-reductive crap. The statistical average. The median take. Answer-shaped, made of nothing you know and nothing you've seen and nothing anyone will remember. Good enough to ship, not good enough to matter.
Want to be middle-of-the-road? Be my guest. You won't raise the round. Just FYI.
The only way out is to refuse the frame that produced all of this, and build the thing that should have been built in the first place.
That is Part 2.