If you use GenAI, you have probably heard about second brains: digital or AI-assisted collections that help people store notes, documents, and ideas so they can find them later. It is a catchy and misleading phrase. Second brains neither think nor understand the world the way humans do. A more descriptive name would be second index, a searchable guide to where information lives.

Humans have been solving versions of this challenge for thousands of years:

  • Libraries need catalogs.
  • Museums need inventories.
  • Businesses need filing systems.
  • Maps need coordinates.
  • Detectives need case files.

Long before computers, people used a technique often called the method of loci or memory palace. They imagined a familiar place, put ideas in specific locations inside it, and then mentally walked through that place when they needed to remember. The basic idea was simple: if you want to find information later, you need a way to organize it today.

The Postal Problem

Most of us rely on one of the largest indexing systems ever built without thinking about it: the postal system.

Imagine a world with only one house. No address is needed. A letter arrives and there is only one place it can go.

Now imagine a village with one hundred houses. Street names appear. Then a city. Then a state or province. Then a country. As more houses are built, the addressing system becomes more detailed: Country. State. City. Neighborhood. Street. Building. Apartment.

Every layer helps narrow the search.

The postal system does not work because every location has a unique name. Quite the opposite. There are thousands of Main Streets. Hundreds of Springfields. Countless Oak Avenues. The system works because enough context is added to remove ambiguity.

  • A letter addressed to "Main Street" is not enough.
  • A letter addressed to "123 Main Street, Springfield, Illinois" is more useful.

The postal system succeeds by creating increasingly precise addresses as the world grows.

Knowledge Has the Same Problem

Information behaves much like mail. At first, a second brain may contain only a few documents. Finding information is easy. Question → Document → Answer.

As more documents arrive, things become more complicated:

  • Different teams use different names for the same thing.
  • A project changes names.
  • A department reorganizes.
  • An acronym gains a new meaning.
  • One group calls something X. Another group called it Y last year. A third group now calls it Z. All three are talking about the same thing.

Humans need to recognize aliases, changed names, and organizational context when a system cannot reliably do so. Just as the postal service must distinguish among many Main Streets, a second brain needs a way to connect different words that point to the same concept. To solve this problem, it builds an index.

  • Topics point to documents.
  • Documents point to sections.
  • Related concepts point to one another.

Over time, the index becomes a map.

How a Second Brain Grows

When a new second brain is created, it has places to put things, but nothing else. Then:

  • A document is added.
  • The system identifies important topics.
  • More documents arrive.
  • New topics appear.
  • Relationships begin to emerge.
  • Eventually, the system creates layers of organization: people, projects, processes, places, products, and policies.

We do not need to know all of these categories up front. They emerge as the collection grows, but not always cleanly. Some categories overlap. Some names change. Some early guesses turn out to be wrong. Just as a city expands by adding roads and neighborhoods, a second brain expands by adding connections and categories. The larger the collection becomes, the more important those connections become.

Following a Question Through the Map

Let's pretend we are building a second brain for a small company. Over time, the company has collected hundreds of documents.

  • Some describe benefits.
  • Some explain vacation policies.
  • Some contain technical instructions.
  • Some document company history.

A tiny piece of that collection might look like this:

Employee Handbook

Benefits Guide

Vacation Policy

New Employee Onboarding

Account Setup Instructions

Password Reset Guide

Network Access Procedures

Now imagine a new employee asks: "How do I get access to company systems?"

A human could read every document. But what if there are hundreds? Or thousands? Humans learned to build, read, and use maps. Today, agentic GenAI, which means AI that can take a series of steps toward a goal, can be designed to do something similar: use a map of the collection instead of guessing from everything at once.

Step 1: Find the Neighborhood

The system first decides which broad area of knowledge seems relevant.

Question

"How do I get access to company systems?"

        │
        ▼

People
Projects
History
Policies
Technology
Benefits

The words "access" and "systems" suggest Technology. So the search moves there.

Step 2: Find the Street

Inside Technology, there are many topics.

Technology

├── Hardware
├── Software
├── Security
├── Account Setup
├── Passwords
└── Networking

The question appears to be about Account Setup. The search narrows again.

Step 3: Find the House

The topic points to several documents.

Account Setup

├── New Employee Onboarding
├── Account Setup Instructions
└── Network Access Procedures

Instead of reading hundreds of documents, the second brain now needs to read only three.

Step 4: Read the Mail

The second brain retrieves those documents and reads the relevant sections.

New Employee Onboarding
    ▼
Complete HR forms

Account Setup Instructions
    ▼
Request company account

Network Access Procedures
    ▼
Wait for manager approval

Now it has enough information to answer.

Step 5: Deliver the Answer

The second brain combines the information and responds:

Complete your HR onboarding forms, request your company account through the onboarding portal, and obtain manager approval before network access is granted.

Notice what happened. The system followed the map instead of reading everything.

Just as the postal system routes a letter through a country, state, city, street, and building, a second brain routes a question through categories, topics, and documents until it reaches the information it needs.

Where Things Break

Many assume second brains fail only because they contain too much information. Our experience suggests that the problem is often about navigation, not just the amount of information.

  • A wrong category can lead to the wrong topic.
  • A wrong topic can lead to the wrong document.
  • A wrong document can lead to a wrong answer.

I saw a small version of this while using Microsoft Copilot to continue a discussion that referenced Andrej Karpathy. In a later response, the system changed the name to “Karpovy” and began treating that altered name as though it referred to the same person and body of work. The answer remained fluent, but the conversation had drifted away from the entity we were discussing.

I corrected it by restoring Karpathy’s full name and explicitly tying the discussion back to the earlier source. The useful repair was not merely fixing one misspelling. It was reestablishing the entity the conversation was about before allowing the system to continue reasoning from it.

I did not trace the system’s internal retrieval process, so I cannot say exactly where the substitution occurred. What I could observe was that the altered name persisted into the answer and had to be corrected before the conversation could proceed reliably.

This taught me that sometimes the reasoning may be fine but the map is pointing in the wrong direction. Other times the reasoning is wrong too. The point is that a bad answer can begin before the answer is even written.

A second brain often fails for the same reason mail gets delivered to the wrong address: the routing system made a mistake.

Humans Still Matter

Modern AI systems can automate much of the work involved in building and maintaining these maps.

  • They can identify topics.
  • They can suggest categories.
  • They can discover relationships.
  • They can even help test whether an index is still working correctly.

Humans still need to decide whether the map truly reflects reality.

The future is unlikely to be humans building indexes by hand, and it is also unlikely to be machines doing everything alone. The most effective systems will likely combine both:

  • AI handles much of the repetitive work.
  • Humans review, refine, and validate the results.

The good news is that nobody needs to start from a blank page anymore.

A collaborative AI can help build and maintain the map. The human's job is making sure the map leads where it claims to lead. And that has always been the hard part.