Twenty-eight days into the second-brain wave, almost every public voice is talking about memory. We have spent eighteen of those days running the four layers around it.

On April 2, Andrej Karpathy posted the idea of a personal LLM Wiki: dump raw sources into a folder and let an LLM build and maintain an interlinked markdown wiki. The post hit 17 million views. Two days later, he published a gist with the details that crossed 5,000 stars and 1,400 forks in 48 hours. I wrote about the idea itself earlier this month; this post is not about the idea, it is about what happens when you run it in production for 18 days for a real business.

The chorus answered immediately. Farza unveiled Farzapedia the same day: an LLM ingesting 2,500 entries from his diary, Apple Notes, and iMessage to build 400 personal Wikipedia-style articles about friends, startups, even favorite anime. Karpathy retweeted it as the canonical implementation. Within 24 hours, Ankur Warikoo coined YBAAS, Your Brain As A Service. By April 13, Aakash Gupta had written the PM playbook with the line that traveled: “the research you did last quarter is gone.” Tiago Forte, the original author of Building a Second Brain, announced the formal end of his 2022 method and launched a new program. Y Combinator put Company Brain on the Summer 2026 RFS. Linas Beliūnas named it as one of the dominant funding themes of the year.

Every public voice is talking about memory.

Memory is table stakes

There is something three weeks of chorus has not mentioned, and it is the part that changes the angle of the problem. Karpathy describes a personal research wiki with immutable sources. Farzapedia is a processed autobiography. Aakash Gupta solves quarter-to-quarter context loss. Tiago Forte rebuilds his personal-productivity method. Even Y Combinator’s “Company Brain”, which is the closest framing in the chorus to actual enterprise operations, is presented as “the missing layer between raw company data and reliable AI automation.” Still a memory framing.

Memory is table stakes. Memory alone is not enough to run a company.

What does not appear in the public conversation: the four layers that turn memory into a working organ. Senses that listen without being asked. Nerves that move information to where execution happens. An immune system that refuses to accept two truths simultaneously. And an encryption layer the entire chorus is avoiding.

What we use the IQ Source brain for in production, today

The brain has been running for 18 days. It is not a research wiki. It is the operational layer of a real services business. Eight concrete uses, no headcount, no client names.

Proposal preparation. Every new opportunity opens against ten or more comparable engagements already living in the brain: pricing precedent, scope decisions, engagement-model fit, what worked in similar industries, what failed. The proposal brief comes out of the brain in minutes, not days, with the source cited. The deck or DOCX comes after, already informed by precedent.

Follow-up tracking across every active engagement. Every call transcript turns into structured pendientes within the same hour. None die in markdown. The brain knows what is pending, what is pending-spec (waiting on client clarification), and what was already handed off to the project tracker, with the commit-hash receipt to prove it.

Client and prospect intelligence. Full relationship history. Every meeting decision, every lesson learned, every upsell signal, every contradiction between what was said in March and what was said in April. We do not re-ask a client a question they already answered.

Meeting ingestion at scale. A call ends, the transcript arrives within the hour, and a structured process reads it for decisions, action items, open questions, contradictions with prior sessions, and changes to the client’s strategic context. Nothing waits for a human to type up notes.

Strategic decision-making. Hiring (when to bring in the first FTE contractor, what to look for, what to pay), pricing (when to hold the line, when to discount), vendor governance (which AI tools to standardize on, which to ban), engagement-model design (when to package as Tech Partner, when as AI Maestro, when as project work). Every decision is recorded with its reasoning, the alternatives considered, and the dissenting view when there was one.

Cross-engagement visibility. A single query surfaces what is pending across every active client, ranked by what is blocking revenue and what is blocking trust. The CEO does not have to remember; the brain remembers.

Daily morning briefing. A review queue surfaces items the brain could not auto-route (a transcript mentioned a company not yet in the brain, a strategic clipping does not fit an existing page, a contradiction needs a human call). Triage takes minutes.

Blog research and pattern recognition. The brain feeds the blog. Every post you have read here in the last month was anchored against precedent the brain already held, including popular voices and the exact prior-cycle analogues from the 1990s onward. The brain is why these posts have receipts.

The four layers that make those uses possible

Senses: ingest automation. Every call transcript enters the brain within the hour. Inbound email signals trigger structured ingestion. Strategy pages from our own blog get auto-extracted and indexed. Daily, weekly, and monthly analytics digests feed the brain without anyone touching a keyboard. The brain receives the world without being asked.

Memory: operational, not aspirational. Encrypted at rest. Credentials, transcripts, legal documents, and pricing precedent never sit in plaintext. Append-only log, never silent overwrites. Every factual claim must cite a source or it does not get written. Citations enforced at write-time, not at audit-time.

Nerves: the handoff bridge. Pendientes from a call transcript do not die in markdown. Three-tier workflow: passive banner when entering a client folder, execute-handoff to push to the project tracker, cross-client status query. Each handed-off item carries pending → pending-spec → handed-off (date, commit hash). The brain acts. It does not just describe.

Immune system: contradiction flagging. When one source contradicts another, the brain refuses to silently pick a winner. It flags the contradiction in a dedicated section and waits for a human. That single rule is what separates an audit trail from a graveyard.

The layer the chorus avoids

Encryption. Last week’s post on the PocketOS incident showed how a vault with three doors open empties a company in 9 seconds. The same logic applies to the enterprise brain.

A personal research wiki holds Wikipedia-grade knowledge. An enterprise brain holds pricing precedent, client transcripts, legal documents, vendor governance taxonomies, hiring signals, strategic positioning. Plaintext is a leak waiting to happen. Encryption at rest with separately-held keys is the minimum bar. And almost nobody in the Karpathy chorus is naming it as a required layer.

The handoff problem in one sentence

Karpathy’s wiki is a destination. An enterprise brain has to be a bridge. Voice in call → structured pendiente → project-tracker entry → commit-hash receipt. Without that bridge, every action item dies in markdown and the brain becomes the most expensive graveyard in the company.

The pattern I have been watching since 1990

I have been doing this for 36 years. Started in 1990, age 15, on a Commodore 64 and a Texas Instruments. I have watched the same knowledge-management cycle repeat five times. Physical filing cabinets in the early nineties. Relational databases in the second half of the decade. CRMs like Salesforce and Siebel from 2000 onward. Notion, Roam, and Obsidian in the last decade. AI brain in this one. Same pattern every cycle: the tool gets adopted before the operational discipline that holds it together. Teams fall in love with the medium and forget that the medium without method produces nothing different from the previous medium.

AI does not change the nature of the problem. It compresses the consequences. A wrong entry in a 1995 filing cabinet hurt one proposal, and the same mistake in a 2005 CRM hurt at most a region. The same wrong entry today, sitting in an AI brain that auto-ingests transcripts, hands action items off to engineers, and cites itself with confidence, scales the lie at machine speed.

Sunday’s post on codebase as moat and the McKinsey piece on the adoption-vs-transformation gap apply here with a twist: the codebase moat and the operational discipline McKinsey names as missing are the same creature seen from two angles. Adopting an AI brain without the four layers around the memory is exactly the category of adoption-without-transformation McKinsey is measuring.

Five questions before approving the next brain

If the executive committee is about to approve the next AI brain, there is a short test worth running before signing.

The first question is about the senses. Most teams answer “we will type into it.” That is the wrong answer. Human ingestion collapses by week three of any project where a client is paying. If the only sensor is a person, what you have is a diary, not a brain.

Then the memory test. The honest version of the question is: are unsourced claims allowed to land? If the wiki accepts a paragraph that has no citation behind it, what you are building is a graveyard where opinions nobody wanted to debate go to die. Mandatory citation at write-time is what separates a useful claim from a rumor.

Encryption is where most committees flinch, because the answer is uncomfortable. What is actually encrypted at rest right now, today? “Nothing” is the most common answer. That answer leaves a leak open for whoever finds it first, whether that turns out to be a journalist or a curious junior at a competitor. The entire Karpathy conversation ignores this, but every security auditor will ask it first when they walk in.

The nerves question is the one nobody wants to answer with a name. Who owns the bridge from the brain to the project tracker? If the response is “the whole team” or “the head of operations,” the operational answer is nobody, and pendientes without a bridge owner die in markdown.

And the immune system. When two sources contradict, what actually happens to the wiki? Silent overwrite is the most expensive option, because it is how you train the brain to lie. Append with flag, or mandatory human resolution, are the only two answers that age well.

If your team cannot answer all five with clarity, the next useful conversation takes two hours. We map the four layers around your memory, answer what is answerable today, and write down what requires intervention. No quote attached. The email is the same as always: info@iqsource.ai.

What we do at IQ Source about this

AI Maestro exists so the four layers around the memory get designed before the brain becomes a load-bearing part of the business. The discovery audit maps the senses, encryption, bridges, and immune system of the client’s current state: where it ingests without being asked and where it does not, what is encrypted and what is not, where there is a bridge to the tracker and where the pendiente lives in a Slack note, what happens today when two sources contradict. Most executive committees discover during the exercise that three of the four layers are missing and the fourth is half-built.

Tech Partner, the other line, applies to software companies whose product lives in the critical zone from day one. For that kind of company, the brain stops being an office tool and becomes part of the deliverable. The client team does not just need to run their own brain; they need to be able to explain to their own enterprise customers why their operational memory system stands up to an audit and to whatever contradiction or encryption review the customer’s security team brings.

Karpathy described a vault. Farza turned the same idea into a personal diary. Tiago Forte read the room and closed his 2022 chapter, and the chorus is now debating which markdown schema wins. Eighteen days of running this in production for a real business points somewhere else. The difference between a vault and a nervous system is not the quality of the archive. It is whether the brain receives the world, remembers with citation, moves information to where execution happens, and refuses to accept two truths at once.