1. The scene
I sit down in a meeting and someone proposes, with conviction, exactly the decision we discarded two years ago. It isn't bad faith — it's the new person in the seat. They run the math anyone would run, arrive at the obvious path. The obvious path we already tested, which cost us, which we recorded somewhere — and no one in the room remembers where.
I half-remember. I don't remember enough to defend the position. I look around and realize I'm the oldest memory in the room — and even I have nowhere to point. The archive is scattered across emails, hallway conversations, minutes no one re-read. We approve the decision again. Eighteen months from now we'll rediscover the same problem. We'll call that learning.
That paragraph isn't a scene, it's a pattern. It repeats in pricing meetings, supplier reviews, process choices. A tribal criterion becomes a rule, a habit, a norm — and no one can reconstruct the reason. When someone asks "why?", the best answer is "we've always done it this way." Each of those meetings costs three ways: the hour spent redoing analysis that existed, the wrong decision when the why is left out, and the trust of new people who realize joining this house means guessing, not learning.
I started writing things down. At first just for me: what I decided, on what basis, against which alternative. It wasn't a method — it was fear of forgetting. Then I started recording before the decision, not after. Within months I had somewhere to point when someone asked "why did we decide this?" — a citable trace, with a date, a context, the discarded alternative.
I started on my own shop floor — the commercial side, where I'm the one who takes the hits. Within three months the same architecture worked for any function that makes recurring decisions. Procurement had it. Operations had it. Finance had it at a different scale. The pain was executive — any function where today's why becomes tomorrow's symptom.
2. The diagnosis
Five pains. I saw them first in the commercial function, then watched the same shape appear in procurement, operations, finance. The vocabulary changed. The geometry was identical.
Critical knowledge lives in heads. "Why did we lose that client in 2023?" — silence, three people glance at each other, someone tries a guess. "Why did we stop working with Manufacturer X?" Could have been quality, a delivery episode, a channel conflict. With no trace, six months later a new buyer requotes the same manufacturer and the house rediscovers the problem the expensive way.
Decisions turn into ghost rules. Tribal pricing logic: this client gets X, that segment accepts Y, below Z we don't go. The new hire grants 12% where the house standard was 6%, and no one notices until the book's margin bleeds three quarters later. A ghost rule only exists while the person carrying it stays.
The report becomes theater. The rep enters the opportunity at the stage they think will land well. "Follow-up pending" gets noted, the screen closes. Next week, same deal, same "follow-up pending." Looking is not deciding, deciding is not recording, recording is not enforcing.
Onboarding is hemorrhage. Six months for a salesperson to read the book without giving away a tell. Four months for a buyer to learn who is reliable on the supplier panel. Three months for an analyst to decode a P&L with proper names and historical exceptions. Nine months for an operations lead. These timelines aren't technical learning. They're about absorbing unwritten context.
The leader becomes a cognitive bottleneck. When critical knowledge lives in heads, someone has to hold the memory of the house. The VP of Sales remembers everything. The senior buyer is the living supplier panel. The CFO is the living register of exceptions. From outside, the house looks efficient. From inside, it depends on four to six heads — each a single point of failure no one dares to name.
3. Why AI alone doesn't fix this
The first reaction of nearly every executive who recognizes these pains is the same. "Then this is a case for implementing AI." The reflex points in the right direction and lands on the wrong object. The problem isn't a lack of artificial intelligence. It's a lack of substrate. And substrate doesn't ship inside any vendor's box.
Karpathy, who ran AI at Tesla from 2017 to 2022 and was a founding member of OpenAI, named the shift most precisely:
"LLMs, not as a chatbot, but the kernel process of a new Operating System."
A kernel is the core of an operating system. It takes instructions, orchestrates processes, returns results. The most sophisticated piece of the machine, and on its own, useless. A kernel with no memory cannot keep what it processed yesterday. A kernel with no file system doesn't know where to find what it needs. That is what the LLM does today, brilliantly, on top of whatever you drop into the context window, and forgets the moment after. The model is the kernel. Your firm doesn't need a better kernel. It needs the operating system around it.
4. The four pillars
Four pieces: memory, policy, decision, cadence. An architecture. Memory without policy is dead storage. Policy without decision is a poster. Decision without cadence is a stray event. Cadence without memory is theater.
4.1. Memory — the single source
Every entity your firm touches (client, supplier, decision, person) has a canonical, living, citable place. One place, not five. When someone asks "what do we know about this client?", is there a single answer anyone can read in three minutes, or eight partial answers in different heads and inboxes? Inside the architecture, the client has a living page: contacts, order history with decisions annotated, context of each negotiation recorded the day of. When the account manager leaves, the next person starts day one with what the predecessor took three years to accumulate.
In AI vocabulary this is called "context engineering": curating the context that feeds the system matters more than choosing the model. The corporate version is building the institutional memory that feeds every decision.
4.2. Policy — rules outside the head
Policy here is not the HR binder on the shelf. It is the set of de facto rules (pricing, discount, supplier qualification, risk escalation, succession) written in prose, versioned, changeable with a trail. Each rule has an author, a date, a reason, and a condition that authorizes its change. The new salesperson reads it on day one and walks into the first pricing conversation with the house behind them, not against them.
In agent vocabulary this is the "system prompt" or "steering" — how you direct the behavior of a system that decides without you in the room. Without explicit policy, every subordinate is an agent without instructions: improvises, with good intent, and each one in a different direction.
4.3. Decision — immutable event with a why
Each relevant decision is an event — date, context, alternatives, evidence, why, criterion for revision. Immutable. You don't rewrite it. You record a new one citing the previous. Today a typical decision happens in a meeting, with three people, and vanishes the minute after. The minutes say "decided: proceed with supplier B." The why doesn't appear. Six months later, when supplier B causes the first headache, no one can reconstruct whether it was a good decision with a bad outcome or a bad one with a predictable result. Inside the architecture, the house learns. Today, it only reacts.
Daniel Kahneman split human thought into two systems: the intuitive, fast one (System 1), and the deliberate, slow one (System 2). AI research has adopted the same vocabulary to describe how current models do almost everything in System 1, and how the next leap is about forcing System 2. Companies have the same problem. Recording a decision is the concrete act of forcing organizational System 2.
4.4. Cadence — the auditable ritual
Cadence is the rhythm by which the house perceives, plans, acts and reflects. Daily, someone notices what changed outside, decides what to move, moves. At week's end, they record what happened, what was learned, what changes next week. Inside the architecture, the four steps have a fixed place — not bureaucracy, ritual. Next week, the same page opens, and the cycle resumes on top of the previous one.
The cycle of perceive, plan, act, reflect is what's now called, in agent vocabulary, the "agentic loop". It's what distinguishes a system that merely responds from one that learns. Without that loop, the agent doesn't evolve. Without that loop, the company doesn't either.
5. What changes — by persona
Architecture, said that way, is still a drawing. What it does for you depends on the chair you sit in.
For the board or shareholder. Today, the only way to assess resilience is a plant tour, three key conversations, and faith. What changes: the house becomes auditable from outside. In valuation, institutional knowledge becomes an asset added, not a risk discounted. Cultural due diligence in an M&A becomes reading trails, not hallway interviews.
For the CEO. You inherited culture, grudges, tacit rules no one told you about in the interview. Delegating today is faith. What changes: the person you hand a situation reads the trail in fifteen minutes and acts inside the criterion.
For the CFO. In November, someone asks why that line item blew up, and you redo the analysis for the third time this year — for the third different person. What changes: each analysis done once becomes a reusable asset. The register of exceptions moves out of your head and into a page with an address. You answer by pointing, not by redoing.
For the COO. The SOP is what the auditor wants. The SOP-with-the-why is what the plant needs when a new lead takes the shift. What changes: each rule gets the why attached. Technical onboarding cuts in half. Human variability drops without more auditing.
For the CCO. The senior rep carries every client's history in their head. When they leave, the house discovers the book the expensive way. What changes: pipeline and account history live outside any one head. Pricing has a defensible trail. The replacement walks in on day one.
For the CPO. The big manufacturer has a dashboard. You have oral memory. What changes: the dashboard moves to your side. Each RFQ learns from the previous one.
For the CHRO. Onboarding is hemorrhage. Every key departure is a cataclysm. What changes: onboarding becomes reading. The context the person was carrying was already in a citable place.
5.5. The byproduct: institutionalized culture
You wanted operational efficiency. You got something larger. Somewhere along the way, someone from outside (a board member, a candidate, a partner) will walk out feeling they understood your firm in a way no one had before. You documented your culture without ever calling it a culture project.
Mission, vision and values in a PDF are theater — not because they're false, but because they describe aspiration, not behavior. The culture workshop documents what the house would like to be on a Monday offsite, not what it decides on a Tuesday when budget is tight. The handbook conveys surface rules, not the tacit criterion by which this house grants a discount, picks a supplier, ends a client.
Real culture is the aggregate of decisions recorded with their why, of policies made explicit, of memory curated, of cadence maintained. In systems that learn from data, behavior emerges from what was recorded during training. In organizations, culture emerges from the traces left when no one was calling it culture. Same physics, different scale. Institutionalized culture outlives the person — the only kind that deserves the name.
6. Where to start
This is not an IT project. It is not an "AI implementation" — AI is the kernel; what's missing is the system around it, and that doesn't come in a box. It is habit reform.
Habit reform doesn't start with the whole company. It starts with one function and one decision-maker. The function — the one that hurts most. The decision-maker — someone willing to record before recording becomes a habit. It doesn't have to be the CEO. It has to be whoever decides something real, every day, and wants to stop carrying the why only in their head.
If this resonated with something you feel but never knew how to name, let's talk. No agenda, no deck, no proposal — just a conversation between operators.
Memory before intelligence. That's the order.