The company that knows itself

Ants returning from food mark their route with scent, and others follow the strongest scent, so routes that keep paying off grow stronger with use. No ant knows where the food is; the trail holds that knowledge, carries it to whichever ant moves next, and lets each result shape what the ants do afterward. Because the scent evaporates, the trail also forgets, and when the food runs out the ants look elsewhere.

The economist Friedrich Hayek found the same pattern in markets, which no one directs and nearly everyone takes part in. In 1945 he argued that the knowledge an economy runs on “never exists in concentrated or integrated form but solely as the dispersed bits of incomplete and frequently contradictory knowledge which all the separate individuals possess.” His answer was the price system, which lets each person act on what others know without having to know it themselves.

Other systems that adapt without a center work the same way. A communications network routes around a failed link because each router keeps a table built from what its neighbors report. An immune system answers a familiar infection faster because some of its cells remember the last encounter. Whatever its form, the trace is what lets the intelligence of the whole exceed anything its parts know.

A company’s trace is its record, everything its work generates, from email and chat to contracts, presentations, and meeting transcripts. Much of what the company has learned is in there somewhere, split across systems that each hold a slice, with none showing how its slice bears on the rest. Unlike a trail, the record doesn’t strengthen what keeps paying off or let go of what no longer holds. Nothing says which plan replaced which, or that a number was later corrected. The company pays for that opacity in repeated mistakes and in decisions made without context.

Even so, companies have become one of the most effective institutions people have built, because people aren’t ants. Each person brings judgment of their own, and over years of working around the opacity, people build up the context the record hides. When they change roles or leave, the documents stay behind, but that context goes with them, including why a decision was made, who the real decision-maker was at a client, and which workaround keeps an old process running.

Some of what people carry can’t be written down. Two decades after Hayek, the scientist and philosopher Michael Polanyi identified this as tacit knowledge: “we can know more than we can tell.” But much of what passes for tacit knowledge in a company could be told, and often has been, in an email, a meeting transcript, or a comment on someone’s draft. People carry it in their heads because the record that holds it is too fragmented to use.

So every company has lived with two problems. Its experience is hidden in systems that don’t connect, and the context people build to work around that leaves when they do. Every competitor had the same problems, so no company fell behind for them.

AI has changed this in two ways. The first is that general intelligence became a utility. Models that read and reason at a level few expected a decade ago are available to any company, and the price of a given level of capability keeps falling.

The economist Tyler Cowen pairs intelligence with what he calls Polanyi knowledge, which covers knowledge of time and place, custom and habit, and much that people can’t put into words. In his account the two are mostly complements, so as intelligence becomes abundant, Polanyi knowledge becomes relatively scarce. So far, he notes, an AI brought into an office can’t work out how it runs, at least not in the human sense. For a company, much of that scarce knowledge is its own experience, both what its people carry and what lies buried in its record.

The second change is that the work itself is moving to machines. AI tools and agents draft, and increasingly they act, sending, updating, and committing on the company’s behalf, often faster than anyone can review. They arrive with none of the context people carry, and they can’t work around the opacity the way people do. An agent acts on whatever part of the record it can reach.

Together, the two changes transform a burden every company shared into a dividing line between them. A company now competes on its own experience, and more of its work is passing to tools that can draw only on the part of that experience they can read.

If the work is moving to models, the obvious question is whether a better model can supply what’s missing or make it unnecessary. Models already reason well over what they’re given, and they’ll keep improving. But what they’re given is a slice of the company’s record, whether a person pastes it in, a search retrieves it, or an agent gathers it from the company’s systems. No single record says whether it still holds or how it bears on the rest, and better reasoning can only guess at what the slice leaves out.

Three ordinary situations illustrate the problem.

The market estimate. An analyst’s rough guess put a new market at $400 million. The number was copied into a strategy memo, a planning deck, and a sales presentation, losing its caveat along the way. A study commissioned later put it at $150 million. An agent preparing board materials finds the larger figure three times and the smaller once, and reports the larger. Three sources outweigh one, except that the three are the same guess.

The ship date. Sales tells a customer a feature will ship in the third quarter. Two months later, engineering moves it to next year and announces the change in its monthly roadmap update. Nothing ties the change to the commitment, which sits in one salesperson’s email. When the customer asks, the account manager’s AI assistant confirms the third quarter, citing that email. Each record is accurate. Together they add up to a broken promise, and the customer is the first to notice.

The channel partnership. In 2022 a company tried selling through a channel partner and failed, for reasons set out in a post-mortem that hasn’t been opened since. In 2026 a new team proposes the same thing, and an AI assistant drafts them a persuasive plan. The assistant would have reasoned well about the post-mortem, but nothing connected the old failure to the new proposal, so it never came up.

In each case the model was capable, and the record held what would have prevented the mistake. Nothing connected it to the work at hand. A colleague who knew the history might have caught the error. An agent working without that context will make the same mistake each time the question comes up.

Training a model on the company’s records doesn’t solve this. A trained model is frozen the moment its training ends, and even retrained without end, it would blend conflicting accounts rather than track which replaced which. Retrieval, citations, timestamps, and an agent’s memory of its own work each solve part of the problem, but none of them, on their own, reliably catch any of the three, because catching them depends on connecting records across systems and over time. That is the work of a living memory, kept outside any model.

A living memory gives a company what the trail gives a colony: one trace that every part of it can act on. It reads across the company’s systems and years, connects what is in them, and keeps all of it current, so that what holds now, what changed and why, what has been promised, and what has been learned are clear to anyone working for the company. People and agents draw on it as they work, instead of each reconstructing a slice of it.

With it, the three situations turn out differently. The agent preparing board materials learns that its three sources are one guess and that a later study replaced it. Before the account manager’s assistant answers the customer, it finds the decision that moved the date. And the 2022 post-mortem reaches the new team before the plan does.

The memory also changes what leaves when people do. Much of the context people carried lived in their heads only because the record was too fragmented to use. Once the record can be used, far less of it leaves with them. What still leaves is what Polanyi meant, the knowledge that was never expressed in any form.

Language models have made reading documents the easy part. Making the company’s record coherent is harder. A customer appears under three names across contracts, invoices, and email, and has to be recognized as one customer. A process changes after an incident, and the reason for the change matters as much as the change itself. When two records disagree, both are kept, because quietly choosing one would hide the disagreement. Everything the memory holds stays tied to its source, and it claims no more than the records show. That is the work Omnisce does.

Models give the memory a voice, and the models will change often. The memory carries the company’s experience from one to the next, and it’s what lets any of them speak for this company instead of for companies in general. It should belong to the company, stay under its control, and follow the company’s own rules about who may see what. Its integrity matters as much as its security. A memory that can be quietly altered, or that confuses what was established with what was assumed, misleads in proportion to the trust placed in it.

Emergent systems learn when results reshape the trace that guides them. Companies learn too, but mostly in pieces. The outcome of a project shows up in one system, while the plan, the assumptions behind it, and the warnings raised along the way sit in others. The team that did the work learns from it, but the rest of the company often can’t, and when the team moves on, much of the lesson goes with it. Post-mortems and reviews close the loop by hand, but they’re costly, so companies save them for the failures that demand them.

With a living memory, the loop can close for the company as a whole. An outcome is connected to the decision behind it and to everything recorded around that decision at the time, so the company can compare what it expected with what happened. The lesson belongs to the company instead of one team, and it’s there the next time similar work begins, whoever or whatever is doing it.

This won’t happen on its own. Results are often ambiguous, causes are hard to establish, and good work sometimes fails anyway, so judging what an outcome means stays with people. Like a trail or a price, the memory informs action without directing it.

That learning compounds, and only for the company that does it. An improvement to a rented model reaches every competitor at once. What a company learns from its own experience reaches only that company and builds on itself year after year, so the distance between companies that learn this way and those that don’t keeps widening.

In 1993 the mathematician and novelist Vernor Vinge observed that science-fiction writers, the ones who tried hardest to imagine what technology would bring, increasingly felt “an opaque wall across the future.” Whether or not that wall has arrived, no one can say with confidence what work, markets, or companies will look like in a few years, and a company can’t plan its way through what it can’t forecast.

Emergent systems manage without forecasts. They act, the results come back through the trace, and the next move follows from them. When the view ahead is blocked, a company is left with the same method, and it works only as well as the company’s memory of what it did and why.

A company’s tools, markets, and people will change, some of them many times. What carries over is its experience, and a company that can draw on it can change course quickly without losing track of what it is or what it has promised.

The companies that know themselves will learn their way through. The rest will keep paying for lessons they’ve already learned.