Source: raw/Every_Level_of_a_Claude_Second_Brain_Explained.md — YouTube video [https://www.youtube.com/watch?v=DTCyvo6cC54], Platform: YouTube. Creator not named on-screen in the transcript; the project name “Herk 2” plus the AI-operating-system 7-day Skool challenge and “four C’s” framework match Nate Herk’s existing wiki coverage, so it is attributed to Nate Herk ^[inferred] (see Open Questions).
A walkthrough that frames the AI “second brain” as five escalating levels of retrieval sophistication — from a CLAUDE.md router up to an always-on autonomous brain OS — and argues you should pick the lowest level that removes your actual pain, not the highest. The creator runs his real “Herk 2” project mostly at level two (an LLM wiki built on Karpathy’s pattern) and treats the whole system as tool-agnostic: “it’s just files and folders,” usable from Claude Code, Codex, or a Hermes agent. The recurring thesis is that your moat is your data, and the hard part is usually getting knowledge out of your head into the system, not retrieving it afterward.
Key Takeaways
- Five levels, each defined by a question; higher is not better. Find the simplest level that fixes a real pain point — “if there’s not pain, then why create more?” The creator deliberately does not run at level five.
- Your moat is your data / IP. The point of a second brain is getting as much out of your head into a system as possible, organized so an AI can recall it without hallucinating or burning time and tokens.
- The real bottleneck is often capture, not retrieval. “Before you blame AI, take a look at your folders and files” — ask whether the system actually holds all the nuance that lives in your head.
- Reverse-engineer from the question. How you will recall data later dictates how you store it now (his basketball-hoop-and-ball analogy: design the ball to fit the hoop).
- The test of a second brain: can your agent find it again, and could you find it again? If no, the routing or folder architecture is wrong.
- Tool-agnostic by design. Everything is markdown files and folders.
CLAUDE.md↔AGENTS.mdare near-identical copies; Claude Code’s auto-memory writesmemory.mdon its own, and Codex just needs a routing rule pointing at it. - One vault can mix levels. A given folder might be level two while another is level four — don’t force a single style (all-wiki, or all-GraphRAG) on the whole project.
- Vector DBs are not magic. Chunk retrieval can miss full context (summarizing only the ~5 chunks it pulled, or naming a “highest sales week” from one chunk while higher weeks sit in chunks it never read). Use markdown when a full-context read matters; use vectors for needle-in-haystack lookups (e.g. “what was rule 17?” out of 1,000 rules).
- Wiki links are not a knowledge graph. Backlinks/“see also” connections lack typed relationships (“endorsed by,” “competitor of,” “works at”) that carry meaning.
- Context vs Connections (two of his “four C’s”: context, connections, capabilities, cadence). Ingest evergreen context; do not ingest fast-changing connections data (Slack threads, emails, customer data) — that becomes noise you delete monthly. Instead give the brain access to fetch it on demand.
The Five Levels
| Level | Question it answers | Mechanism | His verdict |
|---|---|---|---|
| 1 | Find a file/info by exact word or name | CLAUDE.md as router + folders/files | Where everyone starts |
| 2 | Pull everything on a topic together | LLM wiki (Karpathy pattern) + memory.md | Where he runs his whole project |
| 3 | ”I searched different words than I wrote” (meaning) | Semantic / vector search (Qdrant, Pinecone, Supabase) | Use per-folder when routing whiffs |
| 4 | Trace relationship chains (topic X back to topic A) | Knowledge / relationship graph (LightRAG) | Played with it; not daily |
| 5 | Fully autonomous, always-on | Always-on brain OS (GBrain + Hermes) | Deliberately not there yet |
Level 1 — routing. Start with a CLAUDE.md (or AGENTS.md for Codex) that loads automatically and acts like the session’s system prompt. Beyond role and priorities it carries routing rules (“for info about me personally, look in this folder; for Q1 priorities, look in this folder”). If Claude asks for more info when the files already exist, you simply never told it where to look — it won’t scan the whole codebase on its own (and you wouldn’t want it to: wasted time and tokens). Example scaffold: a context/ folder (about-me, stack, a dated decision log) and a projects/ folder (one file or folder per client/project, optionally organized by month). His caveat: there is no proven “standard” layout yet — all that matters is routing that makes sense to you and your AI.
Level 2 — the LLM wiki. Add a wiki built on “the Karpathy LLM wiki” pattern. He keeps separate wikis (YouTube transcripts, meeting transcripts), each auto-created by Claude Code when told to ingest a source, with index → drill-down → read-the-whole-page navigation. Obsidian is only a visualizer of the underlying markdown — nice for visual people, but “all that really matters is can your system grab that and give it to you”; he rarely opens it. CLAUDE.md now routes to wiki + context + projects + decisions + references + memory. Turning on auto-memory (/memory) lets Claude Code maintain memory.md itself; to port to Codex, copy CLAUDE.md → AGENTS.md and add one routing line for the memory file. He runs his entire “Herk 2” project here and hasn’t felt enough pain to move up a level ^[inferred].
Level 3 — semantic search. Chunk a document, embed each chunk so similar meanings cluster in vector space, and search by meaning instead of exact keywords (his “smart lookup” matching “feedback” to evaluation/test-results content the keyword search misses). He demos a Qdrant cluster of images grouping by visual similarity. Key limit: chunk retrieval only sees the chunks it pulled, so summaries and aggregate questions can be wrong — reach for a full markdown read when you need complete context, and reserve vectors for pulling one specific snippet out of a huge pile. You can mix: make just the YouTube-transcript folder a vector store while context/projects/decisions stay markdown.
Level 4 — knowledge / relationship graphs. The most complex and sometimes most expensive tier (open-source options exist). It stores entities (Jordan = person, Acme = company) and typed relationships (Jordan works at Acme; Acme endorsed by PostPilot; PostPilot competitor of Cadently). He shows his own graph in LightRAG, and notes graphs can be more lightweight than wikis (which must read an entire file even when one fact was needed). To feed a graph enough data he uses a “grill me” brainstorm skill (adapted from another creator; name unclear in the captions ^[ambiguous]) that interviews him relentlessly on a topic until it knows everything. He has played with graphs but doesn’t use them day-to-day — his project-based, content-heavy work is served well enough by routing + wikis; a big multi-client CRM would justify a graph.
Level 5 — always-on brain OS. Everything above plus an “always” element that constantly syncs and refreshes memory. He cites GBrain (spelling uncertain in captions ^[ambiguous]), described in the video as created by Garry Tan of Y Combinator and pairing with “GStack,” and notes it fits well bolted onto a Hermes agent (doable in Claude Code but you’d manage the crons yourself). He stays off level five on purpose: auto-ingesting everything risks the “too much context does more harm than good” failure, and he prefers deliberate manual control over what the brain takes in.
Privacy note
The creator flags (in an editor-inserted aside) that routing all of this through Claude means the data goes to Anthropic, which is not private. For client data or anyone uncomfortable with that, he suggests open-source/local models rather than making Claude Code the home of an all-knowing second brain — and says local-AI videos are planned.
Related
- matt-wolfe-ai-second-brain — sibling “second brain” video, but a build tutorial (wiki + journal + CRM on Codex) rather than this level taxonomy
- nate-herk-claude-code-operating-systems-course — the same creator’s AI-operating-system course and “Herk Brain” vault (the project referenced throughout this video)
- karpathy-pattern-third-party-adoption — synthesis of how the wider community is adopting and extending Karpathy’s pattern
- build-llm-wiki-for-business-walkthrough — another third-party build of the wiki pattern, stopping at the wiki layer
- synthadoc — the architecturally-complete end of the spectrum (Python engine + semantic search), i.e. levels 2-3 productized
- headroom-context-compression — directly relevant to the level-five “too much context” caution
- _index — topic root for community implementations of the Karpathy pattern
Try It
- Diagnose by symptom, then pick the lowest level that fixes it:
- Re-explaining your setup / finding things by exact word or filename → Level 1 (tighten
CLAUDE.mdrouting). - 30+ notes and you forget what’s in them → Level 2 (ingest into an LLM wiki with relationships).
- The project whiffs on notes you know exist and routing isn’t working → Level 3 (semantic search, no exact-word reliance).
- You need to follow relationship chains across entities → Level 4 (knowledge graph).
- Running agents offline over tons of data, syncing multiple Hermes agents → Level 5 (always-on OS like GBrain).
- Re-explaining your setup / finding things by exact word or filename → Level 1 (tighten
- Start at level one today: write a
CLAUDE.mdwith explicit “where things live” routing rules plus acontext/andprojects/folder; confirm both you and the agent can retrieve a known file by routing alone. - Make it tool-agnostic: copy
CLAUDE.mdtoAGENTS.md, enable auto-memory somemory.mdself-maintains, and add a one-line routing pointer for Codex. - Let the agent choose the structure per folder: tell your Claude Code agent what a dataset is and how you’ll use it, and ask whether it should be plain markdown or a vector store — don’t force one style on the whole vault.
- Sort new data into context vs connections before ingesting: evergreen, “still useful in a year” context goes in; fast-changing Slack/email/CRM data stays out, with the brain given access to fetch it on demand.
Open Questions
- Creator identity inferred, not stated. The transcript never names the creator or channel. Attribution to Nate Herk rests on the “Herk 2” / “Herku” project name and the AIOS 7-day Skool challenge + four-C’s framing matching nate-herk-claude-code-operating-systems-course (which documents his “Herk Brain” vault and “Herk-2” project). Treat as a cross-reference inference.
- ASR-uncertain proper nouns. “GBrain” / “GStack” (level-five tool, credited to Garry Tan of Y Combinator), the “grill me” skill’s original author, and his quarterly-objectives files (“OTAAs”) are transcribed from auto-captions and may be misspelled. The GBrain claim is reported as stated in the video and is not independently verified here.
- Auto-caption normalizations applied: “Cloud Code”/“cla code” → Claude Code; “Carpathy” → Karpathy; “Enthropic” → Anthropic; “quadrant” → Qdrant; “Pine Cone” → Pinecone; “Superbase” → Supabase; “11 Labs” → ElevenLabs; “codeex”/“codecs” → Codex.