It's always day one

The machine that knows almost everything, but knows no one

Even now, in 2026, I still go through the Barnes & Noble vs. Amazon case with my students. I'm aware it is almost thirty years old, but it illustrates magnificently how the arrival of a disruptive technology can wipe out the competitive advantage of a company with a track record as solid as the B&N of that era, fresh from revolutionizing the publishing world with its superstores. Huge stores, with a selection no neighborhood bookseller could match... and where nobody knew who you were.

To close the case, I like to use a video of Jeff Bezos summing up his management philosophy. Amazon had just announced its acquisition of Zappos, and Bezos decided to introduce himself to his new colleagues with a rather modest recording. Standing in front of a flipchart, he boiled everything he knew down to four ideas.

Obsess over customers. Invent on their behalf. Think long term. And the last one, written in big letters: It's always Day One.

The phrase became one of Amazon's hallmarks. The building where Bezos worked was called Day 1, and when he moved, he took the name with him. More than twenty years after founding the company, he was still reminding shareholders that Day Two led to stasis, followed by irrelevance, then excruciating, painful decline, and finally death. That is why, he concluded, it must always be Day One.

Well, artificial intelligence has managed to take the phrase rather more literally.

Because for a language model, every conversation really is day one. It may know who Ramesses II was, explain quantum mechanics brilliantly and draft a sales contract compliant with Spanish law in seconds. But, with exasperating frequency, it doesn't know who you are, what you are trying to achieve, what decisions you made yesterday or why you ruled out the alternatives.

We have hired the most brilliant advisor in the world and discovered that he suffers from amnesia.

The sage without memories

The problem stems from an understandable confusion. We chat with the system as we would with a colleague down the hall, it answers us with apparent coherence, and we remember telling it something two weeks ago. So we expect it to remember too. But beneath that illusion of continuity, very often there is nothing but an unexpected void.

A language model does not retain the memory of our conversations. It may have told you four days ago that the speech you are preparing for your chairman is on a par with Cicero's first Catilinarian, and remember nothing when you come back to it. Because when it receives a question, it generates its answer from what it learned during training and from whatever information someone has placed in front of it at that moment. That information may include the latest messages in the conversation, company documents or some summary of previous interactions. But the model doesn't remember any of it in the usual sense of the word. It gets told again.

The famous context window works like a desk. While the papers are on it, the model can consult them. When they are removed, they vanish from its world. Making the desk bigger allows the conversation to go on longer, but it doesn't create memory. Confusing a gigantic context window with lasting memory is like expecting a goldfish to remember more because it has a bigger bowl.

The vendors are perfectly aware of this limitation and have been building mechanisms around it for some time. Conversation histories, standing instructions, projects, files, connections to email, and memories that select certain user data to retrieve later. And, of course, a good developer can give their agents external stores in which to note down decisions, results and pending tasks.

But it's not that the model is acquiring memory. It's the application that keeps a notebook and decides what to write in it, when to consult it and which part to show the model.

It sounds like a technical distinction, but it is a formidable business difference.

Finding is not remembering

Companies are trying to solve the problem by connecting artificial intelligence to the data that lives in their management systems. They build RAG systems, index documents, open up email, connect to the CRM or the ERP and let the model consult manuals, minutes and contracts before answering.

All of it undoubtedly useful. But without forgetting that access to a company's information is not the same as knowing it.

A document base can retrieve the minutes recording that the board approved opening a subsidiary in Portugal. But the minutes don't record that Portugal was chosen after ruling out Italy over a regulatory problem, that the CFO agreed grudgingly, that the first candidate to run it lost the team's confidence during the process, and that nobody wants to reopen that discussion.

Companies don't run only on the documents they keep. They also run on a layer that is much harder to capture: discarded decisions, tolerated exceptions, implicit commitments, personal relationships and reasons nobody writes down because everyone in the room knows them.

Exactly: tacit knowledge. Which is also what distinguishes someone who has read a file from someone who has been working with us for years.

A good advisor doesn't add value only because they know a lot. They add it because they remember what we tried last year, who opposed it, where we went wrong and what kind of solution we are actually capable of executing. They may think a decision is technically inferior and still recommend it, because they know the organization that will have to put it into practice. They end up becoming a master of realpolitik.

AI can read all our documents and still show up at the meeting like a newly hired consultant.

And don't expect miracles: giving it memory wouldn't solve this on its own either. The human advisor doesn't remember the Portugal discussion because someone filed it away, but because they were in the room. The problem is not only where to store tacit knowledge, but how to capture it. And I'm afraid the most obvious answer is also the most unsettling: AI will begin to know the organization when it is present in its meetings, its emails and its conversations.

Companies have already seen successes and failures along this road. CRM worked because every sale leaves a record almost without anyone having to write it. The knowledge management of the nineties failed because it asked experts to spend time telling what they knew. AI promises to capture the informal as a by-product, without asking anyone. But it will only be able to store what is said, not what goes unsaid, and the more gets recorded, the more important conversations will move out into the hallway. Because we are human, and we don't feel comfortable under the Eye of Sauron.

The toll of introducing ourselves

This lack of continuity has a cost that rarely shows up in productivity calculations: every time we explain the project again, provide the same background or correct an interpretation we had already corrected, we pay the toll once more. Onboarding doesn't happen once, as it does with an employee. It can happen in every conversation.

And what is merely annoying in personal use can become a structural problem for a company.

One employee spends weeks preparing a proposal with the help of an assistant. Another opens a different conversation and repeats a good part of the work. A third uses another model and gets an incompatible recommendation because the system is unaware of the earlier decisions. Three answers that are reasonable on their own and absurd when viewed together.

The problem becomes more serious when we move from assistants to agents. A chatbot without memory can waste our time. An agent without memory can repeat an action, contradict a previous decision or go back down a path we had already ruled out. Autonomy without continuity doesn't create a digital employee. It creates a succession of new employees who share a name and a password. And burn through tokens by the bucketload.

Which brings us to an even more uncomfortable question: where should that memory reside?

If each employee keeps it, the organization loses part of it when the person changes roles or leaves the company. If the company keeps it, it will have to decide who has the right to read it, correct it and delete it. And, of course, the AI vendor can keep it, which is the most convenient solution and possibly the most dangerous.

Because the memory that makes an assistant useful is also what can make us dependent on it.

The new lock-in

We have spent a few years arguing about which company will have the best model, but perhaps the question will end up mattering far less than it seems. Models improve, their differences narrow and switching models is becoming ever easier. The entire conversation can travel through an API and the new vendor can start answering within minutes.

What doesn't travel as easily is the accumulated relationship.

It is the same lesson Barnes & Noble learned when facing Amazon. Its superstores knew everything about books, yes. But Amazon also knew who the customer was, what they had bought and what might interest them next. And never forgot it. And when selection stopped being scarce, the advantage shifted to memory. Today general knowledge is becoming abundant, and the advantage threatens to shift to the same place.

It's true that tools are starting to appear for exporting the memory of one assistant and importing it into another. But they only move the notebook: a collection of facts, preferences and summaries. They don't transfer the judgment formed while filling it in, nor the intuition about what mattered and what didn't. Receiving someone else's notes is not the same as having been there.

If for three years a system has accumulated how we write, which clients we serve, which risks we accept, whom we consult before deciding and which mistakes we don't want to repeat, switching assistants no longer means replacing a model. It means abandoning someone who knows us in order to introduce ourselves all over again to a stranger.

Memory can thus become the real lock-in mechanism of artificial intelligence. There will be no need to technically prevent customers from leaving. It will be enough to make leaving feel like changing lawyers, doctors or business partners after many years.

Microsoft starts with the advantage of living inside the email, documents and meetings of a large share of companies. Google holds another extraordinary slice of our lives: searches, calendars, emails, files and photos. OpenAI and Anthropic are trying to build that relationship through conversation and daily work with their assistants. They all talk about intelligence, but perhaps they are competing for something much older: being present when things happen, and getting us to entrust them with our memory.

The question for any company is not who will be the Amazon of this story, but how to avoid being its Barnes & Noble: the organization that uses artificial intelligence and lets the memory of its relationship with customers and employees accumulate in someone else's house.

The arrival of Day Two

In Bezos's philosophy, staying in Day One meant keeping the curiosity of a newborn company after accumulating decades of experience. He never wanted Amazon to forget what it had learned, but rather not to become its prisoner.

Language models suffer from the opposite problem. They face each new conversation with dazzling capability, but without a shared biography to lean on. They are always starting over. They have never lived through the previous day with us.

Perhaps the decisive leap for artificial intelligence will come not when a model can answer a harder question, but when it no longer needs us to keep explaining who we are. When it can preserve not just our data, but the story of why we decided what we decided. When it can tell the important from the anecdotal, lets us correct its memories and is able to carry them over if we switch vendors.

In an innovative company it may be healthy to think it is always Day One. In a relationship, however, everything truly valuable begins on Day Two.

The question is who will hold on to the memory when we finally manage to get there.