Your company learns at machine speed. It remembers almost none of it.

For a long time you could dismiss this as a curiosity: a few enthusiasts chatting with a bot. That framing is dead. Anthropic’s analysis of 1.2 million Claude Cowork sessions across more than 600,000 organizations found that software development accounts for just 8.7% of usage. Much of the rest is the operational core of companies: business process and operations alone is 33.4%, the single largest category, alongside content creation, DevOps, research, data analysis, document processing and sales operations[1]. Anthropic calls it “the work around the work”. We would call it something else: it is the exact layer of a company where tacit knowledge lives.

OpenAI, meanwhile, reports over one million business customers and ChatGPT Enterprise seats growing 9x year over year[2]. LLM traffic measured on a single leading model router grew by an order of magnitude in roughly a year, into the quadrillions of tokens[3].

Put plainly: a substantial share of your company’s daily reasoning now happens inside model interactions. The interactions grew by an order of magnitude in about a year. The organizational record of them is, in most companies, zero.

In 2025, MIT’s NANDA project published the number every conference speaker has repeated since: 95%. It is usually misquoted. The report’s actual claim is that 95% of organizations surveyed were getting zero measurable P&L return from their GenAI investments within the observation window, and that just 5% of custom enterprise AI tools made it to production. The authors themselves flag the figures as “directionally accurate”, based on 52 interviews and 153 survey responses, with a six-month window that may understate success[4]. Treat the number with care.

What almost nobody repeats is the report’s diagnosis. It is worth quoting exactly:

The report’s interviewees say it in plainer words. A corporate lawyer: the tool “doesn’t retain knowledge of client preferences or learn from previous edits. It repeats the same mistakes and requires extensive context input for each session.” The most cited barriers among users: “it doesn’t learn from our feedback” and “too much manual context required each time”[4]. Read that against the adoption data above and the picture is uncomfortable. Companies are not failing at AI because the models are weak. The models got good faster than anyone expected. Companies are failing because every session starts from institutional amnesia, and every correction an employee makes is a gift the organization throws away at logout.

It is tempting to think the vendors will simply fix this with bigger context windows or a memory toggle. The technical evidence says otherwise. First, context does not behave like memory. The “Lost in the Middle” study showed that model performance degrades sharply when relevant information sits in the middle of a long context; in the worst cases, giving the model the right document buried among many others performed worse than giving it nothing at all[5]. Chroma’s evaluation of 18 frontier models reached the same conclusion from another angle: performance grows increasingly unreliable as input length grows, even on trivially simple tasks[6]. A longer window is a bigger desk, not a filing system.

Second, persistence across sessions does not exist by default. The entire research line on agent memory, from MemGPT onward, starts from the same premise: anything outside the window is gone unless an external system deliberately captures and re-injects it[7]. When researchers gave web agents the ability to store and reuse workflows they had already figured out, success rates improved by 24.6% and 51.1% in relative terms on standard benchmarks[8]. That delta is a measurement of exactly how much value evaporates when nothing is written down.

Third, even the people building frontier agents work around this, manually. Anthropic’s own engineering guidance recommends “structured note-taking”: the agent writes notes outside the context window, because critical context “would otherwise be lost” across long tasks[9]. Vendor memory features exist precisely because, as OpenAI put it when launching ChatGPT’s memory, remembering across chats “saves you from having to repeat information”[10]. These are real improvements, and they are scoped to a person and a product. Your account remembers you. The company remembers nothing. That distinction is the whole problem, so let us make it explicit.

When a senior analyst spends twenty minutes teaching an AI assistant the company’s revenue recognition edge cases, three things are true in most organizations today. The correction improves that one conversation. It may improve that one person’s future conversations, if their memory feature caught it. It will never reach the colleague two desks away who will make the same mistake tomorrow, nor the agent that colleague runs, nor the auditor who in two years will ask why two systems answered the same question differently.

Organizations have always bled knowledge this way, before AI existed. McKinsey estimated interaction workers spend nearly 20% of their week looking for internal information or tracking down colleagues who hold it[11]. Panopto’s workplace study put a number on the departure side: 42% of institutional knowledge is unique to the individual holding it, and knowledge workers waste over five hours a week waiting for or recreating knowledge that already exists somewhere in the company[12]. Both studies have caveats, and both predate LLMs. That is precisely the point. AI did not create the organizational forgetting problem. It multiplied the volume of knowledge being created while leaving the retention rate at zero.

And there is a darker layer. Most of this is not even happening in accounts the company can see. Microsoft’s Work Trend Index found 78% of AI users bring their own tools to work[13]. Cyberhaven’s telemetry, based on actual usage rather than surveys, found that 73.8% of ChatGPT accounts used at work are non-corporate accounts, and that the share of sensitive corporate data pasted into AI tools more than doubled in a year, from 10.7% to 27.4%[14]. IBM’s 2025 breach report attaches a cost: organizations with high levels of shadow AI absorbed an average of $670,000 in additional breach costs[15]. The knowledge is not just unretained. It is accumulating, unmanaged, in places the organization cannot audit at all. We covered the governance side of this in AI governance for SMEs.

The regulatory clock. The EU AI Act’s general applicability arrives on 2 August 2026, bringing transparency and high-risk obligations, with deployer duties phasing in from there[16]. The EDPB has already published a risk methodology specifically for LLM systems[17]. We will not pretend this regulation is about organizational memory; it is not. But it makes one question unavoidable for any European company running AI in real workflows: what did your systems know, where did that knowledge come from, and who approved it? A company whose operational knowledge lives in a million deleted sessions has no answer. On why data sovereignty matters here too, we wrote a dedicated piece.

The competitive clock is more interesting. Y Combinator’s Summer 2026 Request for Startups includes an entry called, literally, “Company Brain”. Tom Blomfield’s framing: “The biggest blocker to AI automation of companies is no longer the models... Now the blocker is the domain knowledge”, and companies need “a system that pulls knowledge out of all these fragmented sources, structures it, keeps it current”, because “AI agents can’t operate like that” on knowledge scattered across heads, threads and tickets[18]. When Y Combinator declares a missing primitive, funded teams follow. The MIT report found the same dynamic from the buyer side: executives’ top demands from vendors were systems that learn from feedback (66%) and retain context (63%), and one CIO put the endgame bluntly: “once we’ve invested time in training a system to understand our workflows, the switching costs become prohibitive”[4].

Memory, in other words, is becoming the moat. The only open question is whether it will be your moat or your vendor’s. It is the same underlying question as who owns the company brain.

We have spent months on this problem, first inside our own company, then with others. We do not have a finished answer. We have become confident about the properties any real answer must have, and we offer them here as a checklist, mostly because we would like to be challenged on them.

The tensions are the point: the teams that resolve them, in whatever form, will define how companies remember in the agent era. As for where to start measuring where you stand, we formalized it in the Company Brain Maturity Model: the L0-L4 scale of how a company’s operating knowledge lives.

Here is the uncomfortable arithmetic to run on your own organization. Take the number of AI interactions your teams had last month, sanctioned or not. Estimate the fraction that contained a correction, a decision, a rule, an exception: knowledge your company paid salary to produce. Now ask what percentage of it your organization could retrieve today.

For almost every company we have talked to, the honest answer is a rounding error above zero. Your company is learning at machine speed. It is remembering at approximately zero. That gap compounds daily, it is invisible on any dashboard, and if the MIT diagnosis is right, it is what separates the few organizations extracting real value from AI from everyone else.

We are researching how organizations are dealing with this today: workarounds, internal tools, policies, resignation. If you run AI or agents in production and this problem sounds familiar, we would genuinely like to compare notes. Write to us. The worst outcome is a good conversation.