The Forgetting Machine: Why Companies Cannot Remember What Works

The Forgetting Machine: Why Companies Cannot Remember What Works AI August 28, 2026 17 min /ai/the-forgetting-machine/ The US government once forgot how to make a nuclear weapons material. NASA managers ignored their own lessons-learned system. Companies store everything and remember almost nothing, and AI agents are about to industrialize the problem. Inside the quiet race to build the layer that lets organizations actually learn.

In the year 2000, the United States government discovered that it had forgotten how to make a component of its own nuclear warheads.

The material is called Fogbank. It sits between the stages of a thermonuclear weapon, and its exact nature is still classified. It was manufactured at a facility in Oak Ridge, Tennessee, between 1975 and 1989. Then the facility was shut down, the people who made it retired or moved on, and almost nobody wrote down how the process actually worked. When the government needed to refurbish its warheads a decade later, it found the recipe was gone. Not the documents. The knowledge.

According to the Government Accountability Office, the effort to relearn how to make Fogbank cost $69M and pushed the warhead program a year behind schedule.¹ And buried in that story is a detail that should haunt every executive who believes their company's knowledge is safely written down somewhere. When engineers rebuilt the process from the old records, the new material failed. It turned out the original had worked because of an impurity that the old process introduced by accident. Nobody had ever documented it, because nobody knew it mattered.²³ The organization had known something that none of its documents knew.

The Saturn V story is similar, and similarly misunderstood. The internet loves to claim NASA "lost the blueprints" to the rocket that went to the moon. It did not. The drawings survived. Rocketdyne even ran a formal knowledge retention program, recording interviews with the engineers who built the F-1 engines. What dissolved was everything else: the supplier base, the tooling, and above all the people. When the company surveyed itself in 1992, it found just 248 active employees with F-1 experience left, plus 76 retirees it could call back.² One study estimated that restarting production would cost $315M in 1991 dollars, not because the design was missing but because the living system that produced it was.

Here is the uncomfortable question. These are the most consequential engineering organizations in American history, working on the most important machines ever built, with formal programs whose entire purpose was to preserve knowledge. And they still forgot. What are the odds that your company, with its wiki nobody updates and its CRM full of empty fields, is doing better?

Companies have never recorded more. Every call can be transcribed, every email archived, every document versioned, every dashboard preserved forever. Storage is effectively free and effectively infinite. And yet organizations know less about themselves than they think, because storage and memory are different things. Storage keeps what happened. Memory keeps what it meant.

The gap between the two has three causes. They behave less like trends than like laws: every organization anyone has measured obeys all three.

The first law: knowledge lives in people, and people leave. The median American worker has been with their employer for 3.9 years as of the latest official count, the lowest figure since 2002.³ For workers aged 25 to 34, it is 2.7 years. During the churn peak of 2022, more than 50 million Americans quit their jobs in a single year. Meanwhile the most experienced cohort is walking out the other door: more than four million Americans are turning 65 every year through 2027, the largest retirement surge in history. The research organization APQC surveyed companies about this in 2025 and found that 92% do not consistently capture knowledge from employees who are about to retire.⁴ Not "fail to capture everything." Do not consistently even try.

The second law: the knowledge that matters most is concentrated in shockingly few heads. In a survey of over a thousand employees at large US firms, 42% of institutional knowledge was reported to be unique to the individual holding it⁵: when that person leaves, nobody else can do what they did. Software has a blunt name for this, the bus factor, and researchers have measured it. A study of 133 of the most popular open-source projects on GitHub found that about a third depended so heavily on a single developer that losing that one person would cripple the project. Two thirds would be crippled by losing two.⁶ These are the most transparent, most documented engineering projects in the world. Corporate knowledge is almost certainly more concentrated, not less.

The third law: even the knowledge that stays decays. This is the strangest one. Economists have measured how fast organizations forget, and the numbers are startling. Studying Lockheed's aircraft production line, the economist Lance Benkard found that roughly 39% of a plant's accumulated production experience effectively evaporated every year; labor hours per plane actually went back up mid-program as the line changed and people rotated.⁷ A study of ambulance services found something even more useful: when researchers separated the causes of organizational forgetting, employee turnover accounted for roughly twice as much knowledge loss as individuals simply getting rusty.⁸ Organizations do not mainly forget the way people do, by fading. They forget the way buckets do, by leaking.

None of this is a new observation, and that is exactly what makes it interesting. Corporate America has already run a full-scale, decade-long, multi-billion-dollar attempt to solve this problem. It was called knowledge management, it peaked around the year 2000, and it failed so completely that the phrase itself became faintly embarrassing.

The pitch will sound familiar. In 1999, the analyst firm IDC warned that Fortune 500 companies would lose $12B that year to "knowledge management inefficiencies," and forecast that spending on KM systems would explode. Half of all large US companies planned to install one.²⁴ The knowledge base, the intranet, the corporate wiki, the lessons-learned database: an entire industry was built on the theory that if you gave employees a place to write things down, the organization would remember.

It did not work. A Computerworld investigation in 2000, right at the peak, quoted a research director at British Telecom putting the failure rate of KM initiatives at around 70%.⁹ Bain's long-running survey of management tools ranked knowledge management among the lowest-satisfaction tools it tracked, year after year, even in years when most surveyed firms used it.²⁵ And the most damning evidence comes from the organization with the most at stake. In 2001, the GAO surveyed NASA's own program and project managers about the agency's lessons-learned system, built after Challenger precisely so that hard lessons would never be lost.

A quarter of NASA's managers did not know the system existed. 43% had not submitted a single lesson in two years. Fewer than one in five had ever applied a lesson found in it.¹⁰ Two years after that survey, the Columbia accident investigation board concluded that the organizational causes of Columbia were substantially the same as the causes of Challenger, seventeen years earlier.¹¹ The lessons had been learned, written down, stored, and never used.

The pattern repeats wherever anyone has measured it. The Project Management Institute has found that organizations waste roughly a tenth of every project dollar on poor performance, year after year, for the same recurring reasons, while only about a third of project-office directors rate their organization as good at feeding lessons from failed strategies back into planning.²⁶ The management scholar Paul Nutt spent two decades tracking more than 400 real corporate decisions and found that about half failed, meaning they were abandoned or never fully used within two years.¹² Even the most quantitative companies on earth have this problem. Google, LinkedIn, and Microsoft each run more than 20,000 controlled experiments a year, and about two thirds of tested ideas fail to improve the metric they target. That failure rate is fine; that is what experiments are for. What is not fine is that, as the field's leading textbook laments, most organizations keep no institutional memory of their experiments at all.¹³ They discover the same negative results over and over, at industrial scale.

Why did knowledge management fail? Look at the autopsies and one cause towers over the rest. The systems captured documents, but a document is an answer with the reasoning removed. The wiki page says what the policy is, not why, not what was tried first, not what happened when the previous policy met reality. Writing things down was a tax paid by the busiest people for the benefit of people they would never meet, so mostly they did not pay it. And crucially, nothing ever connected a decision to what happened next. The lesson database and the results lived in different systems, maintained by different teams, joined by nothing. An organization cannot learn from experience if its experience is never assembled in one place.

A document is an answer with the reasoning removed. Companies filed millions of answers and kept none of the reasons, and then wondered why they kept relearning the same lessons.

For thirty years, corporate forgetting was a chronic disease. Annoying, expensive, survivable. What turns it acute is what companies are doing right now: hiring millions of new workers who cannot remember anything at all.

AI agents are arriving in the enterprise at a pace that makes every previous software wave look slow. Microsoft reported that in a single quarter, customers created over one million custom agents through its Copilot tools, up 130% from the quarter before.²⁸ Gartner's latest research places agentic AI at the very peak of its hype cycle and notes something remarkable: only 17% of organizations have deployed agents so far, but more than 60% expect to within two years, the most aggressive adoption curve for any emerging technology it measures.¹⁴

Here is the problem. Today's agents are, by construction, amnesiacs. A language model does not update itself when it works for you. Each session starts blank; whatever the agent figured out yesterday is gone unless someone stuffs it back into the prompt. Researchers building the first benchmark for lifelong agent learning put it flatly: current agents "remain stateless and unable to accumulate or transfer knowledge over time."¹⁵ Andrej Karpathy has compared working with them to working with a colleague who has anterograde amnesia, brilliant in the moment and unable to form new memories. The writer Dwarkesh Patel, in a widely-cited essay on why this matters more than raw intelligence, offered the best analogy I have seen: it is like trying to teach a student the saxophone by having each new student read written notes about the previous student's mistakes. "The reason humans are so useful," he wrote, "is not mainly their raw intelligence. It's their ability to build up context, interrogate their own failures, and pick up small improvements."¹⁶

This is not an abstract complaint. Ask anyone who uses a coding agent daily: it will confidently repeat the exact approach you rejected last week. The startup Mem0, which sells memory infrastructure for agents, describes the demand for its product in exactly these terms: users watching assistants re-suggest solutions that were already turned down. And the most-quoted enterprise AI study of the past year, MIT's survey of generative AI pilots, named this as the root cause of its famous finding that 95% of pilots showed no measurable profit impact. The report's own diagnosis got far less attention than the headline number: most generative AI systems "do not retain feedback, adapt to context, or improve over time." The learning gap, the authors called it.²⁷

Put the two halves together. Companies were already forgetting machines: leaking knowledge through turnover, filing answers without reasons, never connecting decisions to results. Now they are wiring in a workforce of agents that forgets by design and executes at machine speed. An organization that learns nothing per action, multiplied by a thousand times more actions, does not become smarter. It becomes confidently, rapidly, expensively wrong. It repeats yesterday's mistakes faster than any workforce in history.

The people building this technology know all of this, which is why, over the past eighteen months, a race has quietly started to build the missing memory.

The intellectual starting gun was fired in April 2025 by two of the most decorated researchers in the field. David Silver, who led AlphaGo, and Richard Sutton, the father of reinforcement learning, published an essay called "Welcome to the Era of Experience." Their argument: the knowledge that can be extracted from human-written data is approaching its limit, and the next generation of AI will improve mainly by learning from its own experience, from streams of actions and their consequences rather than static text.¹⁷ Whatever you think of the timeline, the framing stuck. Within a year it was cited in venture theses, product launches, and lab roadmaps alike.

The labs moved first on the consumer side. OpenAI gave ChatGPT memory in 2024, extended it to reference all your past conversations in 2025, and in June 2026 shipped a feature that consolidates what it knows about you in a background process between sessions. OpenAI named it Dreaming, which says the quiet part out loud: the models are being given sleep²². Anthropic went the other direction and shipped memory to businesses first: project-scoped memory for teams and enterprises arrived in late 2025. Microsoft added memory and personalization to Copilot. Each of these is real progress, and each is also, so far, personal memory: your assistant remembers you. The company remembers nothing.

The startups are attacking the deeper layer. A wave of "memory infrastructure" companies has formed: Letta, founded by the Berkeley researchers behind MemGPT, raised $10M declaring memory the key to self-improving AI. Mem0 raised $24M and now processes nearly 200 million memory operations a quarter for other people's agents. Cognee raised $7.5M to build what it calls an enterprise memory layer. These are small checks by 2026 standards, rounding errors in a year when AI took more than 60% of all venture funding. But watch what the big companies are paying for the adjacent thing: the context that agents need. ServiceNow paid $2.85B for Moveworks, the largest acquisition in its history.²⁰ Workday paid $1.1B for Sana, a knowledge platform.²¹ Glean, which builds a permission-aware graph of everything a company knows, was last valued at $7.2B.

The investors, meanwhile, have talked themselves into a remarkable reversal. In 2019, Andreessen Horowitz published a famous essay arguing that data moats were mostly empty promises. By 2023, Sequoia was conceding that the data flywheel it once predicted had not materialized and that moats would come from customers and workflows. The 2025 and 2026 essays have converged somewhere new. Bessemer's State of AI report says it plainly: "For AI application founders, context and memory may be the new moats."¹⁹ Sequoia's Sonya Huang now tells founders that with strong post-training and online learning, an application company can beat the frontier labs inside its own domain. And Salesforce's AI research group, in an essay on self-improving agents, produced the sentence that I think will define the next five years of enterprise software: "A model is something you rent. A loop belongs to you, and it's something that compounds."¹⁸

Follow the logic of that sentence. Models are becoming interchangeable commodities; I have written before about how the value is migrating away from them. Data, the last decade's favorite moat, turns out to be inert on its own: a warehouse of exhaust. What compounds is neither the model nor the data but the loop: a system that observes what the organization does, ties it to what happened next, and feeds the difference back into the next decision. The industry keeps circling the same missing primitive from different directions. Gartner's newest hype cycle even gave it an early analyst name, tucked among the agent platforms: context graphs.

So what would it actually take for a company to remember? Work backward from the failures and the requirements almost write themselves.

First, it has to capture decisions with their reasoning attached, not just their conclusions. The one great counterexample to corporate amnesia is instructive here: Amazon's memo culture. By banning slide decks and forcing every major proposal into a six-page written argument, Bezos accidentally built something no knowledge base ever achieved, an archive of why. Every significant decision at Amazon has a reasoned document attached to it. That is the unit that matters: not the document, the decision, with its context riding along.

Second, and this is the step every previous system skipped, it has to connect each decision to what happened next. This is the join that knowledge management never made: the lesson lived in one system and the result lived in another, so the organization could never check its own work. Close that loop and something changes qualitatively. Repetition becomes evidence. If the same move works nine times in similar situations, that is no longer an anecdote in a departed employee's head; it is a pattern the organization itself holds, with a track record attached, and with the counter-examples preserved too.

Third, it cannot depend on people doing homework. The KM era died on the capture tax: asking the busiest people to stop and write. This is the part that has genuinely changed. Work now leaves a machine-readable trail by default: the calls are transcribed, the messages are logged, the tickets and commits and contracts are all in systems, and open protocols increasingly let software read across them. For the first time, capturing what happened and why is an engineering problem instead of a compliance program.

Fourth, the memory has to be usable at the moment of the next decision, by whoever is making it, human or agent. A memory you have to go visit is a library; nobody went to the last one. A memory that shows up inside the work, that tells the new account executive what the founder noticed in a hundred similar deals, or warns the agent that this exact fix failed twice before, is something else entirely. It is the difference between a company that stores its experience and a company that compounds it.

Enterprise software has been organized around the same layers for thirty years: systems of record to store what is true, and systems of engagement and action to do the work. The layer that is missing, the one all these converging efforts are groping toward, is the one that learns: the system that gets a little harder to compete with every single time the company acts. The firms that get this will stop treating experience as exhaust and start treating it as the asset. Everyone else will keep paying the Fogbank tax, one departed expert and one amnesiac agent at a time.

Notice what kind of argument this is. It does not depend on a forecast about model capability or a guess about adoption. It follows from five statements that are each independently measured: knowledge lives in people; people leave; what stays decays; documents do not carry reasons; agents multiply actions without accumulating anything. Reject the conclusion and you must reject one of the premises, and every premise has decades of data behind it. The missing learning layer is not a product idea. It is a consequence.

If the premises above hold, and they are measurements, not opinions, two things follow. I will make them concrete enough to be wrong.

• Prediction one: the learning layer becomes a named category of enterprise software. The way "system of record" and "CRM" became fixtures of the architecture diagram, a system that links a company's decisions to their results and feeds the patterns back into daily work will become a standard, budgeted layer of the stack. The early signals are already visible: Gartner naming context graphs, Bessemer calling context and memory the new moats, every agent platform sprouting a memory feature. The test: within a few years, analyst firms will rank vendors in this category by name, and enterprises will run formal evaluations for it the way they once did for data warehouses. If agent adoption keeps growing while memory remains a per-user personalization feature, I was wrong.
• Prediction two: improvement rate replaces launch-day capability as the way AI systems are judged. Today every buyer asks how good the model is. As models converge and commoditize, the question that separates winners will become: how much better is this system at my company after six months than it was on day one, and can I keep that improvement if I leave? Contracts, benchmarks, and pricing will reorganize around measured improvement on the buyer's own work, the way some support agents are already paid per resolved case rather than per seat. The test: the standard enterprise AI evaluation will include a learning-curve clause, and "it does not get better here" will become the standard reason for churn. If, instead, buyers are still choosing systems on static benchmark scores years from now, I was wrong.

The deepest irony of the AI era may be this: we spent decades worrying that machines would replace human workers, and the first thing the machines actually inherited from us was our amnesia. The companies that fix that, that finally close the loop between what they do and what they learn, will not just have better software. They will be the first organizations in history that truly remember. Everyone else has already forgotten this article.