The memory patch.

Ask companies what they use AI for and you hear about production: drafting content, summarizing documents, writing code. Ask why their AI initiatives die and you hear about something else. When a research group at MIT's Media Lab interviewed enterprises about their generative AI programs last summer, the barrier that kept surfacing was not infrastructure, regulation, or talent. It was that the tools "do not retain feedback, adapt to context, or improve over time." One corporate lawyer in the study, equipped with an expensive specialized contract tool, preferred a plain chatbot for first drafts and trusted neither with anything serious: "it doesn't retain knowledge of client preferences or learn from previous edits. It repeats the same mistakes."

I think the pattern has a plain explanation. Most of what companies actually need from workplace AI is memory, and above all memory of decisions, individual and organizational. Large language models were never built to remember anything. The industry knows this, which is why an entire ecosystem of retrieval pipelines, vector databases, and "memory" features now exists to compensate. The question worth asking is the uncomfortable one: is that ecosystem a reasonable bridge, or are we patching a defect in the underlying technology with clumsy workarounds when it needs to be fixed from first principles?

My answer is that it is both, because "memory" is two different problems wearing one name, and only one of them belongs inside the model.

A note on evidence before I lean on it. The MIT report was preliminary, its headline numbers were disputed, and the group behind it promotes an agent architecture that happens to match its conclusions, so I cite its interviews rather than its percentages; the interviews match what I hear from teams directly. The technical claims are not contested; they are in the founding papers. And the argument has a falsifier: if a frontier lab ships a model that genuinely consolidates experience across sessions, half of this essay becomes history. The other half, I will argue, survives even that.

Start with what the models are. The founding objective of the GPT line, unchanged since 2018, is to predict the next token of text drawn from an enormous frozen corpus. Training compresses that corpus into the model's weights; Andrej Karpathy calls what remains "a hazy recollection" of what it read. After training, the weights freeze. Inference is stateless: the model re-reads the entire conversation on every turn and retains nothing when the session ends. The one native way to give a model new knowledge, more training, fails in a well-documented way: teaching a network new information degrades what it already knew, a failure mode called catastrophic forgetting. Karpathy's description is the honest one: an LLM is "a coworker with anterograde amnesia."

So the industry did what you do for a brilliant colleague who cannot form memories: it surrounded them with notes. Retrieval pipelines fetch documents into the prompt. Vector databases index everything ever written. The "memory" features in the major assistants are dossiers, summaries of past chats pasted invisibly into new ones. None of this touches the weights, because touching the weights breaks the model. The patches help, and they perform like patches: on benchmarks built from long, multi-session conversations, accuracy falls off steeply even for long-context models, and the dedicated memory products still miss a meaningful share of what they are asked to recall, by their own published numbers.

Whether that counts as clumsy depends on which memory problem you mean, and here is where I think the debate goes wrong: it treats memory as one problem.

The first problem is experiential. A useful employee gets better at the job: remembers what was tried, what failed, what you meant last time, and stops repeating mistakes. That is what the lawyer was asking for, and it is what the scaffolding fakes least convincingly, because a dossier of notes is not skill. This problem really does live in the model, the patches really are patches, and the first-principles fix is an open research problem. Ilya Sutskever argues that scaling alone has run its course and the next step is models that learn on the job; prototype architectures already update their own weights at test time. I expect this to be solved in the models eventually, with no confidence about when.

The second problem is institutional, and it is the one businesses feel most. Call it decision memory: what was decided, by whom, on what reasoning, which alternatives were rejected, and what would justify reopening the question. This is the substance organizations run on. Management theorists defined organizational memory decades ago as stored history "that can be brought to bear on present decisions." Workers feel its absence more than they want help creating: in Microsoft's large annual survey of workers, most said they lose time hunting for information, and finding things outranked creating things on the list of what they wanted AI to do. Onboarding is mostly downloading decision memory from people who were there; when they leave, it walks out the door with them.

Here is the uncomfortable part for the first-principles instinct: decision memory does not belong in a model's weights, and it never did. It has never lived inside any single brain. Institutions solved memory with records, not neurology, because institutional memory must be things a neural network is not: inspectable, correctable, auditable, shareable, and owned by the organization rather than by one participant. Suppose you could fine-tune your company's decisions into an agent. You would have recreated the problem you were solving: memory trapped in one head, impossible to audit, corrected only by surgery, and lost the day you retire the model. A first-principles fix for decision memory is external by design. The scaffolding is not wrong about that. It is just aimed at the wrong scale: dossiers about one user's preferences, retrieval over documents that happen to exist, context assembled per session and thrown away. It patches the individual's problem and ignores the organization's.

A concrete version. In March, a company decides to stop selling to hospitals. The reasoning lives in one heated meeting, a spreadsheet, and a direct-message thread between three people. In September, someone asks an agent to draft next year's go-to-market plan. The agent is excellent: it reads everything it can reach, reasons carefully, writes fluently, and includes hospitals, because nothing it can reach contains the decision. The model failed at nothing it was built for. It produced relevant text from the information that existed; the thing that mattered was never recorded anywhere at all. Two gaps stack: the model cannot remember, and the organization never wrote it down. No architecture fixes the second gap. A model that learns from experience can only learn from the experience it is present for.

The strongest objection to all this says context windows will dissolve the problem: frontier models already pass needle-in-a-haystack tests over millions of tokens of context with near-perfect recall. The trend is real and the tests flatter it; recall degrades sharply when the task requires meaning rather than string matching, and a context window empties when the session ends. Bigger working memory is genuinely useful, and it leaves the question untouched: what fills it, and from where?

So, to answer the question I opened with. Where the flaw is real, the fix from first principles is coming from the labs, on their schedule, and the notes-and-dossiers era will look in retrospect like the workaround it is. But for the memory businesses need most, the model was never the right place, and waiting for a better model is waiting for the wrong shipment. The first-principles fix for decision memory is an environment: a shared place where people and agents work together, where decisions get recorded as a side effect of being made, visible to the next participant who needs them, human or not. Documentation as a separate act of discipline has failed in every generation of knowledge management that tried it. The record has to be a byproduct of the work, which means the work has to happen somewhere that can hold it.

A company is, more than anything else, the set of decisions it does not have to make twice.

Organizations have always been memory machines built from forgetful parts. Every generation of employees rebuilds context the last one carried out, and we treat the loss as a law of nature. We are now adding participants whose memory could in principle be complete, shared, and permanent, and we are deploying them in the one configuration that guarantees they remember nothing. The models were built to write, and they were patched to cope. Whether anything remembers is not a property of the technology. It is a choice about where we let the work happen, and it is still ours to make.