AI Has a Dementia Problem: Why AI Keeps Starting Over

I worked at Shopify for a brief stretch. There is one moment from that time I keep coming back to.

I had a problem to solve. I pulled seven people into a room — everyone connected to the work stream, everyone with relevant context in their head. We had a clear solution in an hour. Everyone in the room was confident in it. The kind of clean, scalable answer that should ship the next day.

It did not ship for almost three weeks.

What happened in between is the part I want you to see clearly. When we sought approval to roll out the solution, we were told we couldn't solve it that way. The people above the work — the ones who needed to sign off — could not explain why. They could only tell me it had to work that way. Somewhere, some time ago, someone made a decision, and as a result, it worked this way. Nobody could unpack the reasoning.

I needed to make a good decision. So I went on a treasure hunt. I went to the person whose role I had taken over. Same answer: that is how it is done. I went to twelve more people over the next two weeks. Same answer, different mouths. The team was fully aligned on how things worked. None of them could have a rational discussion about why.

Nobody could justify why we had to keep doing it that way.

We shipped the new solution. The original work had taken seven people one hour. The work to get permission to actually do it had consumed two-and-a-half weeks of calendar time and the better part of a working week of additional human attention across twenty people.

This was one small decision within an organization of thousands of people working on hundreds of things simultaneously.

Bain & Company has spent a decade quantifying what they call organizational drag — all the coordination overhead, approval cycles, and inherited practices that fill calendars without producing decisions. In Time, Talent, Energy, Mankins and Garton documented that the average company loses 21% of its productive capacity to it. The decision rights live above the operational context. The institutional memory of why something is done a particular way decays one promotion cycle at a time.

What I lived through at Shopify was the human-and-process layer breaking down. The pattern does not change when you add AI. The same problem compounds exponentially. Across more decisions. By smaller teams. With less visibility. Every day.

What broke at Shopify is what breaks in every organization: the decisions, the work the decisions land on, and the outcomes the organization needs aren't being held in the same space. That disconnect is a memory failure — the institutional recall, the documentation, the record of why we do what we do, all breaking down together. The organization loses its mind one decision at a time.

This is a pattern I saw repeat in an unexpected place. In a stack of note cards my father keeps beside his living room chair.

The cards

My father is sitting in his chair watching television. A commercial comes on for a memory supplement. He picks up a pen. He grabs a four-by-six blue-lined note card from the stack on the side table next to his chair. He writes down the supplement's name and website so he can look into it later. He adds the card to the top of the stack he has already written.

This is one of hundreds of cards. They are not just next to his chair. They are by his bed. They are taped to the bathroom mirror. They are on the backs of cupboard doors. They are in places he trusts he will see them again.

My father has dementia. He was diagnosed five years ago. I have watched the slow loss of continuity, the strategies the brain develops to compensate, and the moments when the strategies stop working.

The cards are not new. He has used note cards his whole career. He was a world-renowned microbial ecologist who specialized in biofilms. He worked alongside founders of the field like Bill Costerton. His work appeared in places like the Gaia Circular, where he wrote "The origin of communities," and in the Annual Review of Microbiology. He was part of a lineage of thinkers who challenged whether evolution can be understood through individual selection alone — Lynn Margulis, whose husband was Carl Sagan, was among them. I grew up hearing stories of my parents going over to her house for dinner.

The cards were how he captured the ephemeral: a brief epiphany that arrived when he was away from his microscope, not focused on the work. Shower thoughts. He would write the thought down, store it in a place he trusted he would come back to, and return to it later in the lab. The card preserved the thinking at the moment it had arrived. Meeting notes do the same work for any team: compress the discussion into a synthesized form a future reader can decompress without recomposing from scratch. Without the card, the thinking slips away. The cards were a thinker's tool.

And somewhere along the way, my Dad's thinking tools became a crutch.

The moment the cards became a crutch

I was sitting in the doctor's office with my mother in the room where they review brain scans. The doctor put up two images. One was a healthy brain. The other was my father's. The healthy brain looked whole. My father's looked like Swiss cheese. There were holes where there should have been tissue. There was less brain to work with.

After his diagnosis, the cards started doing something else. Same instrument, different function. After seeing the brain scans, I decided to start looking through his cards to see what he was writing down.

They no longer captured ideas at the boundaries of human understanding and scientific exploration. They were no longer notes for the book he was trying to write. I would go through them and find the same thing written again and again. The same supplement, the same name, the same appointment. They captured what he could not afford to lose. The names. The days. The appointments. The things he had decided yesterday. His stacks of cards, once filled with brilliant insights, were filling up with small moments he was trying to hold on to.

He was trying to keep track of a world he could no longer hold together in his mind. He couldn't remember conversations. He couldn't piece together his short-term memory. He couldn't recall it. It wasn't being written properly. Or stored. He could no longer rely on his own memory. He couldn't pull the context of his life together. There was nothing he could look up. But he kept writing the cards, even when they could not help him remember.

He lost his grip on reality. One card at a time.

We would have the same conversation six times in the same day. We would unpack the reasoning and walk through the motions. These were big decisions, the kind that would have required him to be a part of them in a normal context. He knew they were important. He knew he should be involved. But he couldn't remember having the conversations or coming to the decisions with us. He would get angry when we told him we had already talked about it. Upset that he wasn't part of it.

It was frustrating. We were working with someone who couldn't hold context in their working memory.

Same pattern, different system

A few years later, I was sitting at a desk, pasting context into Claude for the fourth time that week. I was building a project that needed continuity — what I had decided yesterday, who was involved, what the shape of the thinking looked like. The tool did not have that. So I was reintroducing it, every time, by hand.

I recognized something with a coldness I did not expect.

The tool was not broken. It was incomplete. There was an architecture for holding context that was not there yet. The ability to recall what we had decided. To capture what we had explored. To document the reasoning. To weave the threads together across sessions and across days. The technology was not yet capable of holding that space. So I was holding it. I was one of the cards.

I was making my own cards too. Markdown files holding decisions, processes, and information. Context windows refilled by hand. MCP connectors stitching systems together so they could share what each knew. CLIs scripted to recover what the LLM had forgotten between sessions. Multi-agent workflows where one agent carried memory for another. Each one a workaround for an architecture that was not there yet.

If you do not recognize those terms, you are in good company. Even now, with AI everywhere, the share of people working at this depth — building with agents, wiring connectors, managing context — is far below the share of people using it at a surface level.

We rely on these systems because our own context limits are not enough. The scope of the spaces we work in. The complexity of the systems we build. The skills and experience and perspective held by each person connected to the work. All of it together is a distributed information system. Each piece. Each system. Each person. Every one is a card.

What we are still missing is the architecture to reassemble that context on demand. To hold it in our minds when we need it. To rebuild it when we do not have it. To use it the moment it matters.

Computational systems have been converging on the function of the human brain for centuries. You can see it in the language of every era of computing. Memory. Attention. Context. Reasoning. Each one a new attempt to make machines do what people do unconsciously. Even in Claude itself, where Anthropic has shipped Dreams, a feature that reorganizes Claude's memory between sessions, the way a sleeping mind consolidates the day.

The convergence is still happening. As these systems become less deterministic and begin to explore the emergent and the ephemeral, the same challenges and frustrations we know from working with people will show up in the synthetic thinking machines we are just beginning to build.

This was the same pattern I experienced communicating with my dad. It became more pronounced as his dementia progressed. In a different system now. The architecture could not hold what we needed it to hold. The human picked up what fell. As the context expanded, the architectures holding it broke down. The same way my father's cards grew until they covered every surface. In the same way, his context window kept shrinking.

Except for one difference. My father's case is biological. The architecture is damaged and is not coming back. The AI case is structural. The architecture is missing and is being built.

Every organization is already doing this

Once you see the cards, you see them everywhere.

Open the notepad someone on your team carries everywhere, with handwritten information. Open the whiteboard photo from last week's ideation session that nobody has come back to — if anyone thought to take the photo before it was erased. Open the Google Doc, Notion page, or SharePoint folder your team pastes into the AI tool every time someone starts a new chat. Open the meeting-transcript pipeline your most senior engineer wired through an MCP connector that the rest of the team isn't quite sure how to use yet. The information lives everywhere. Most days, we are cobbling it together. To avoid rework. To move faster. To make better decisions.

These are note cards. They are the externalization of context that the system underneath cannot hold on its own. The system that cannot hold context, in this case, is your organization — the collective system made of people, processes, technology, and time.

We have made a mess of these cards. We rarely come back to most of them. So we reinvent the wheel. We build what another part of the business has already built and does not know about. We pay the cost of finding it again. The cost of reassembling work and thinking that has already been done. The cost of decisions blocked by people who do not have the context but spout dogma based on faulty assumptions or tradition. It is tradition asserted in the void where reasoning should have been. The story I opened this article with was that cost, made visible. All of it is how we compensate for organizational memory failure. Or we wake up every day and silently accept the consequences.

Organizations have been doing this for over a century. Standard operating procedures, written processes, training manuals, decision frameworks. These are what organizations build when the human inside the role cannot carry forward enough context to make the role transferable. When somebody leaves and somebody new starts, the cards are what continuity runs through.

Every leader I have worked with, at every scale of company, has a stack like this. They have built primer documents and pasted them into AI tools. They have written onboarding pages. They have started decision logs. They have asked the team to document why we did it this way. The work has been continuous, because the underlying problem is continuous: the system the work runs through cannot hold its own context, reassemble it in a meaningful way for people to use, or improve upon it when it is not working for us. And the people inside it have to carry it.

The thing the cards have always done — at every scale, in every era — is hold context in a place someone trusts they will come back to. The cards on my father's bathroom mirror. The internal knowledge system your team keeps. The system prompts you to paste into a fresh AI session. Same move. Different substrate.

What breaks when the cards aren't enough

More cards is not the same as the cards being enough. Four things start to happen when the cards run out of their intended utility.

The first is improvisation. The cards have become a crutch, and when the crutch is not enough, the system fills the gap by looking for signals it can use to infer an answer. My father has lost track of where he is in the day. He sleeps six times per day, and every time he gets up he thinks it is morning. At dawn or dusk, he can no longer tell which is which, so he reads the quality of the light and guesses. AI does this when it cannot remember and produces what we have learned to call a hallucination: an invented citation, an invented person, an invented decision. An organization does this when nobody can tell you why something is done a particular way, so the team produces an explanation that fits the pattern and feels right.

The second is judgment failure. Ask the system for a recommendation and you will get one, delivered with confidence, unmoored from the context that would have made the recommendation good. My father has done this with money decisions. He would go out and spend thousands on things we would later have to return, which signaled that something wasn't working right in his brain. Judgment is one of the first things dementia takes. The AI has done this with strategy. The organization has done this when the people above the work make calls without the context the people doing the work hold.

The third is unawareness. The system cannot tell you it has lost track. From inside the experience, nothing is missing. My father is not aware that his memory capacity is smaller than it was. The AI cannot tell you what context it has dropped to keep working with you in the conversation. The organization cannot tell you it is operating without the context that would justify its decisions; "it has always been done that way" is the local version of "from inside, nothing is missing."

The gap is invisible until you find yourself having the same conversation or correcting the same mistakes.

The fourth is complexity. The system grows faster than our ability to map and integrate it. My father has this: the same thing written in five different stacks, contradictions across the years, no clear sense of when each card was made or what was true at the time. AI has this when context windows hold redundant or contradictory snippets and the model cannot tell which version is current. The organization has this when documentation grows faster than the team can maintain or integrate it, and the surface area of what must be managed compounds beyond what any single person can hold.

The Shopify moment I opened with was these four failure modes at organizational scale, running concurrently, all the time, on hundreds of decisions. The hour I spent in a room with seven people who carried the context produced one outcome. The two-and-a-half weeks consuming twenty people's time produced the same outcome, at much higher cost.

That is what the cards prevent when they are working. That is what unprotected gaps produce when the cards are no longer enough.

Where the metaphor breaks

The comparison between dementia and AI has held across every failure mode. But this is where the two cases diverge. They are moving in opposite directions.

In dementia, we do not know how to reverse cognitive decline. The architecture is following a course of entropy. The cards cover less and less ground each year. We are not going to fix the architecture. The cards are the practice that lets us hold what we still can.

AI is the inverse. The system started in a disabled state, and we are developing its capabilities. The architecture is not damaged. It is missing. It is being built.

Multiple groups of engineers and researchers are working on the same memory and context engineering problem organizations have navigated since Frederick Taylor introduced the standard operating procedure in 1911. At Anthropic. At OpenAI. At Google. In published research everywhere. People working on mechanical computation have navigated it for over two centuries. SOPs were an early form of externalizing specialized context into shared systems — the same architecture we are now building into AI. The question is not whether the architecture is buildable. It is.

Boston Consulting Group's October 2024 survey of a thousand C-level executives found that 74% reported struggling to scale value out of their AI investments. The failure is not in the algorithms. It is at the layer where people and processes meet the technology — roughly 70% of the gap.

The deeper question is what we are building toward.

For the long-form history of how we got here, Walter Isaacson's The Innovators is a good place to start. What I want to point out are three moments when the pattern of human-and-machine collaboration came into view.

In 1945, Vannevar Bush published "As We May Think" in The Atlantic Monthly. He described a hypothetical machine he called the memex — "an enlarged intimate supplement to his memory." The memex would let an individual extend their thinking by recording associative trails and coming back to them later. Bush was writing a year before the first stored-program computer ran. The vision was already there.

In 1968 Doug Engelbart took the stage at a computing conference in San Francisco. About a thousand people in the room. He gave a demonstration that was later called the Mother of All Demos. In ninety minutes Engelbart introduced the mouse, hypertext, the windowing interface, video conferencing, and a real-time collaborative editor. He was not showing what computers could do instead of humans. He was showing what humans could do with computers as collaborators.

In 1998 Garry Kasparov played the first match of what he called advanced chess, against another grandmaster, with both players using chess engines as collaborators. Computers paired with humans, not against them. The format eventually evolved into freestyle chess. The teams were called centaurs. Half-human, half-machine. By the early 2000s, centaur teams were producing chess at a level neither pure humans nor pure machines could match.

I have to acknowledge what happened next. By around 2017, AI engines surpassed centaur teams in chess specifically. Pure AI now wins. But the lesson from chess-freestyle survives. Kasparov made the case two decades ago: weak human plus machine plus a better process beats a strong human plus machine plus an inferior process. The differentiator was not the human's capability or the machine's capability. It was the process. The way the two worked together.

Chess is a fully-defined game. Finite states. Perfect information. A clear scoring rule. Most of what leaders actually do is not chess. The decisions that matter at human scale require intuition, context, judgment, lived experience. Exactly the work that happens in the room when seven people who hold the context come together and solve a problem in an hour. Working as a team, not on one. Not in silos. Not in lanes. Not in positions.

There is one more thread in this lineage I want to bring in. My father's research, and the field of microbial ecology he helped shape, demonstrated that bacterial communities are computational systems. Not as metaphor. Communities of microbes solve problems together, share state through chemical signals, and adapt collectively in ways no individual cell can. Computation is not something humans invented and tried to teach to machines. It is what biology has been doing in community for billions of years. The architecture we are now building into AI is the architecture biology already discovered.

This is the work of cocreation. Humans with each other, humans with machines, and now machines with other machines — synthetic systems retrieving and sharing what each one knows. All of it in process together, holding context and reasoning collectively, building what no single participant could build alone. When the process is built well, it produces the best outcomes. That is where we are headed. Bush and Engelbart and Kasparov saw the shape of it from where they stood. My father saw it from a different angle, looking inside the communities of microbes he studied. For fifteen years, I have seen it from inside the rooms where teams come together to do what no single member could do alone.

What to think about this week

I want to come back to the room I opened this article with — the seven people I pulled together to solve a problem at Shopify in an hour.

That hour was the first cocreative move. I knew the practice by another name then: facilitative leadership. What AI engineers now call context engineering, I was doing at human-team scale, before they had a name for it. The move was there: pulling together the people who held the context, building a shared place where the context could be made visible and worked on, producing an answer the system underneath could carry forward. The work was designing a conversation to unpack the collective context, intelligence, and wisdom of the people, then converging on a solution that solved more problems instead of solving a symptom and distributing new ones across the system and people who depend on it every day.

The two-and-a-half weeks of interviews and approvals that followed were the cost of not having that shared place ready before the problem arrived. Twenty people's worth of time and three weeks of calendar were paid out trying to reconstruct context that had never been externalized — context that, like my father's short-term memory under dementia, had been lost forever. The system did not have a card to come back to.

For fifteen years I have spent most of my professional life facilitating that first move at human-team scale. The work is the same regardless of the team or the problem: getting information out of people's heads into a shared space they trust they can reference, come back to, and use to improve the quality of the decisions they make, the speed of their execution, and the rework they avoid. Building the systems that hold context inside teams and organizations so the systems can do what people alone cannot.

That is the work AI rollouts need now. The same move, with AI now in the room.

The leaders building these tools are at the leading edge of a practice that has always been the bottleneck. Context engineering on the AI side. Facilitative leadership on the human side. Same practice. Different substrate. The leaders who already do this with their teams are equipped to do it well with AI now in the room.

If you are running an organization and using AI right now, this is what I would have you sit with this week.

Notice the work you are already doing. You are holding context for intelligences that cannot hold it on their own. Your team. Your tools. The AI in the room.

The work I have been describing is not new to you. It is the work you have been doing all along, without naming it. Sitting with that, this week, is the move.

The next question is whether the organization you are building has dedicated space — actual time, actual people, actual systems — for that work to happen well. Not as overhead. Not as documentation theater. As the leadership move that determines whether the cocreation arc your team is on the front edge of becomes value-creating, or becomes another two-and-a-half-week delay on hundreds of decisions a year.

The cost of doing that work in the time you've set aside for it is small. The cost of discovering you needed to do it in twenty interviews three weeks too late is what we just walked through.

The cards have always been kept by someone holding context the system itself could not. A caregiver for a family member who's ill. A senior team member. A leader. From now on, they are kept together: by the team, with each other, and increasingly with the tools we are building to hold the context with us.

This is the leadership work I believe is most underdeveloped. Stewardship. Of the systems we are in. Of the work we do. Of the culture of our teams and companies. Of the communities we are a part of. And the weight of that stewardship is not for one person to carry. It is distributed across the people and the systems we share the work with.

Caldwell Leadership exists to help organizations build that practice. That is what this work is.