All video notes
Video notes · Notes and PKM

Every Level of a Claude Second Brain Explained

Video by Nate Herk | AI Automation on YouTube · 31 min · uploaded June 2026

Notes and analysis by BrainTube, written from the video. Not affiliated with Nate Herk | AI Automation. Watch on YouTube

In short

Nate Herk lays out five levels of a Claude Code knowledge setup, from a routed CLAUDE.md over markdown folders up to semantic search, knowledge graphs and an always-on agent, and argues you should pick the lowest level that removes a real pain. That advice is sound, and his warning about chunked vector search losing whole-document context is the most useful part. The levels mix up storage and retrieval, though, and the claim that plain files work with any AI tool only holds for tools that can read your disk.

The argument

The video's argument is that a personal knowledge base succeeds or fails on one question: can you and your agent find a thing again? To answer it, you design backwards from how the data will be used, the way you would shape a ball to fit the hoop. He then ranks five approaches by the question each one answers: exact names and words, everything on one topic, matches by meaning, chains of relationships, and fully automatic upkeep. Higher is not better. He runs almost all of his own project at level two, and he says different folders in one project can sit at different levels. The rest of the video argues that the hard part is getting your knowledge out of your head and into files, not the retrieval technology.

Key ideas, with timestamps

  1. 2:12

    Design for the question first

    Before any level, he asks how the data will be recalled, because that decides how it should be stored. His test for the whole system is simple: could the agent find it again, and could you? If not, the routing or folder structure is wrong.

  2. 4:23

    Level one: CLAUDE.md as a router

    The project file that loads every session carries routing rules: personal context lives in one folder, quarter priorities in another. Claude Code will not search the whole project on its own, so a missing route is why it asks you to re-explain things. His example folders are context, a decision log that the agent appends to with dates, and projects.

  3. 8:13

    Level two: an LLM wiki

    He uses the Karpathy-style LLM wiki for his YouTube and meeting transcripts. Claude Code ingests a transcript and writes concept, comparison and technique pages with index files. Obsidian at 9:20 is only a viewer for those markdown files, and he admits he rarely opens it.

  4. 10:58

    Moving between agents

    To keep the setup tool-agnostic, he copies CLAUDE.md to AGENTS.md for Codex and tells Codex to read the memory.md file that Claude Code's auto memory keeps updated. At 22:59 he adds that CLAUDE.md can just reference AGENTS.md instead of duplicating it.

  5. 12:36

    Wiki links are not a knowledge graph

    Wiki links work like 'see also' backlinks: they say two pages are connected, not how. A knowledge graph stores typed relations such as 'works at' or 'competitor of'. The wiki gets a similar effect only because the agent follows the trail and reads full pages.

  6. 15:53

    Where vector search fails

    Level three is semantic search. His warning: ask for a summary of a March 5 meeting stored as 20 chunks and the agent summarises only the few chunks it retrieved. Ask which week had the highest sales and it may answer week 6 from one chunk when week 14 was higher. Whole-document questions want a full markdown file; vectors suit lookups like 'what was rule 17' in a list of 1,000.

  7. 23:33

    Level four: typed relationships

    His sample graph has Jordan works at Acme and Postpilot is a competitor of Cadently, and he shows his real data in LightRAG. He does not use graphs day to day. To feed one, he uses a customised 'Grill Me' skill, which he credits to Matt Pocock, that interviews him until a topic is fully written down.

  8. 25:11

    Level five and what to keep out

    Level five is an always-on, self-syncing setup such as Garry Tan's Gbrain, which he is testing with Hermes Agent rather than Claude Code because of the cron work. He prefers control over what gets ingested. At 26:49 he splits data into evergreen context, which belongs in the knowledge base, and changing connections such as Slack threads and email, which the agent should fetch live instead.

What it gets right

The strongest advice is the least exciting: start with routing. Most failed setups are not missing a vector database. They are missing a line that says where a kind of information lives. Treating CLAUDE.md as an index rather than a dumping ground also keeps the always-loaded context small.

The chunking warning at 15:53 is correct and rarely said in videos that sell vector search. Top-k retrieval returns fragments, so questions that need a whole document or a whole table (summaries, maximums, totals) come back confidently wrong. His split between evergreen context and fast-changing connections is also a good rule. Ingesting every Slack thread makes retrieval worse, not better.

The mid-video privacy note at 21:19 is honest. Anything processed through a hosted model is sent to that provider, and client data deserves a decision, not a default.

What it leaves out

  • The levels mix two separate choices. How you store knowledge (files, wiki pages, graph) and how you retrieve it (exact match, meaning, relations) are different axes. Hybrid search, which combines keyword and vector matching, is now a common default and covers levels one and three at once. He names hybrid search and re-ranking only in passing at 18:04.
  • The chunking problem has standard fixes. Store a summary per document, keep metadata such as dates, and let retrieval return the parent document rather than the chunk. He mentions metadata briefly, but the takeaway viewers get is 'use markdown instead', which does not scale past what an agent can read in one pass.
  • Graph extraction is not free or exact. LLM-built graphs cost tokens to build, can invent or miss relations, and need re-running as data changes. He calls graphs expensive but does not say how to check one.
  • 'Any AI model' is narrower than it sounds. Files and folders work for agents with file access, such as Claude Code, Codex or Hermes. A chat app on your phone cannot read your laptop's folders. Copying CLAUDE.md to AGENTS.md also creates two files that drift unless one imports the other.
  • No measurement. There is no test for whether a level actually answers better. Even ten saved questions with known answers would tell you when to move up.

How this applies to memory across AI tools

His core test, whether the agent can find it again, gets harder once you use more than one assistant. A folder of markdown serves the agents that run on your machine. It does not serve the chat window you use on another device, and each tool's own memory file stays inside that tool.

The usual fix is to put the knowledge behind a server that any client can query, which is what one memory across ChatGPT, Claude and Cursor describes. BrainTube takes that route for the evergreen layer he describes: it saves YouTube videos with their transcripts, articles, PDFs and podcasts into one compiled knowledge base and exposes it as an MCP server, so Claude, ChatGPT, Cursor and other MCP clients can search it. His transcript wiki is exactly this kind of material. Keep his rule either way: evergreen sources go in, fast-changing threads are fetched live. For the keyword versus meaning trade-off itself, see semantic search vs keyword search.

Who should watch it

Watch it if you already use Claude Code or Codex on a folder of notes and are deciding whether to add a wiki, a vector index or a graph. The most useful stretches are 4:23 to 5:29 on routing, 15:53 to 17:31 on where vector search fails, and 26:49 to 28:26 on what to ingest versus fetch live. The closing checklist at 28:57 maps symptoms to levels. Skip it if you want build instructions for any single level: it is an overview with demo folders, not a tutorial, and it includes two plugs for the creator's free community.

Frequently asked

Do I need a vector database for my AI notes?
Usually not at first. A routing file that tells the agent which folder holds which kind of information solves most 'it can't find it' problems. Vector search helps when you have a lot of text and ask narrow questions by meaning, but it returns fragments, so it handles summaries and totals badly.
What is the difference between an LLM wiki and a knowledge graph?
An LLM wiki is a set of markdown pages the model writes and links together, like backlinks: it records that pages are related. A knowledge graph stores entities and typed relations, such as a person who works at a company. Graphs let an agent follow relation chains without reading every full page, but they cost more to build and maintain.
Can I use the same CLAUDE.md setup with Codex?
Yes, for the parts that are plain files. Codex reads AGENTS.md, so either copy the file or have one file reference the other to avoid drift. Claude Code's auto memory writes to its own memory file, so point Codex at that file in your instructions.

Keep videos like this one searchable

BrainTube saves YouTube videos with their transcripts into one compiled knowledge base that Claude, ChatGPT, Cursor and other MCP clients can search.

Start free

Go deeper

More video notes

Are you the creator and want this page changed or removed? Contact us and we will act within 48 hours.