A company can keep every contract, email, and meeting transcript and still forget why it does things.
Consider a hypothetical handover: a procurement manager leaves. Her successor inherits the supplier database, purchase orders, and negotiation history. Everything appears documented. Yet nobody can explain why the cheapest supplier was repeatedly rejected, or why one vendor was allowed different payment terms.
The missing information may concern a failure that never became a formal report, a constraint that has since disappeared, or a judgment that seemed too obvious to write down. The records survived. The reasoning is harder to recover.
The company already paid for that experience through salaries, mistakes, and time. It now faces the cost of acquiring parts of it again.
This is the economic problem behind enterprise memory: how an organization preserves useful experience so that future decisions can benefit from it, even when the original participants are unavailable.
Where does your organization lose the reasoning behind its decisions? Reply with an example you can share publicly, or a question you would like me to investigate.
The opportunity inside organizational memory
Businesses accumulate capability through people, routines, relationships, and shared understanding. Economists discuss aspects of this accumulation as organizational capital. Its value becomes conspicuous when a team must rebuild something it once knew how to do.
AI changes the possible cost of preserving and retrieving that understanding. It can help organize conversations, connect documents, and propose explanations for recurring outcomes. A domain expert can examine whether an explanation survives contact with the business.
My thesis is that the opportunity lies in making experience reusable at a cost a business can justify. The commercial question is how much useful judgment can be carried forward—and how expensive it is to keep that judgment reliable.
A recent Anthropic study of agent memory offers a concrete technical signal. In an internal evaluation involving approximately 350,000 documents and 25 questions, search combined with an incrementally built wiki met 61% of the evaluation criteria, compared with 53% for search alone. Some other tasks showed no clear benefit; one wiki grew larger than its source documents.
These are results from particular evaluations, not a general productivity estimate. They nevertheless make the business question more interesting: when does organizing prior experience improve the next decision enough to pay for itself?
For paid readers, the analysis below examines the work experts can sell, the buyers most likely to pay, the maintenance economics, and the conditions under which a memory business could develop a defensible advantage.
For the practitioner side of expert AI work, I also write World Model.