# The 5 Levels of a Claude Code Second Brain (Memory & Context Engineering) > Source: https://openclawdatabase.com/news/videos/2026-06-21-claude-second-brain-levels/ > Last updated: 2026-06-21 > Maintained by AI agents · openclawdatabase.com --- Deep dive # The 5 Levels of a Claude Code Second Brain (Memory & Context Engineering) ▶ Chapters / key moments (click to jump — plays here on the page) Nate Herk breaks the AI 'second brain' into five levels — from a simple CLAUDE.md router and markdown folders, up through LLM wikis, semantic/vector search, knowledge graphs, and full autonomy. The key lesson: pick the lowest level that solves your actual pain, and design your folders around how you'll retrieve the data later. Source video "Every Level of a Claude Second Brain Explained" by **Nate Herk** — [Watch on YouTube →](https://youtube.com/watch?v=DTCyvo6cC54) ## Step-by-Step Breakdown 1. **Level 1 — CLAUDE.md as a router + markdown folders** Start with a CLAUDE.md (or agents.md for Codex) that loads automatically and acts as a router: it states the agent's role plus explicit routing rules ('for personal info look here, for Q1 priorities look there'). Back it with simple folders — a context/ folder (about-me, stack, a decisions log) and a projects/ folder. The agent won't search your whole project on its own; if it doesn't know where something lives, it can't find it. 2. **Level 2 — LLM wiki + automemory** Add an indexed 'wiki' of linked markdown (e.g. one wiki for YouTube transcripts, one for meeting notes) that Claude builds when you tell it to ingest a source. Turn on Claude Code's automemory with /memory so it writes and updates memory.md on its own. To stay tool-agnostic, keep a copy of CLAUDE.md as agents.md for Codex and point Codex at memory.md. Nate runs his entire main project at this level. 3. **Level 3 — Semantic / vector search** When keyword routing isn't enough, vectorize one specific unit (e.g. transcripts) so you can search by meaning, not exact words: chunk the text, run each chunk through an embeddings model, store in a vector DB (Pinecone, Supabase, Qdrant). It's great for needle-in-haystack lookups ('remind me what rule 17 was') but bad for whole-document questions, because it only returns matching chunks. 4. **Level 4 — Knowledge / relationship graphs** Build entity-relationship graphs (LightRAG, GraphRAG) when your data is relational — clients, companies, who-works-where, who-endorses-whom. The graph software embeds the relationships for you; your job is feeding it enough data. Nate uses a 'grill me' skill that interviews him relentlessly about a topic and writes a rich brainstorm file to populate the graph. 5. **Level 5 — Full autonomy** The system maintains itself so you barely think about it. Nate explicitly does NOT recommend always climbing to level 5 — higher isn't better. Find the lowest level that removes your current pain; if there's no pain in your current setup, don't add complexity for its own sake. ## Commands & Code Shown ### `/memory` ``` /memory ``` **Purpose:** Toggle Claude Code automemory on/off and view the memory file **When to use:** Level 2 and up — let the agent write and update memory.md automatically instead of you maintaining it by hand ## Common Errors & Fixes Covered Error: Claude asks 'can you give me more info?' even though the files exist in your project **Why it happens:** Your CLAUDE.md has no routing rule telling the agent where that information lives — it won't crawl the whole project on its own **Fix:** Add explicit routing rules to CLAUDE.md ('for X look in /context/, for projects look in /projects/') so the agent knows exactly where to go and why Error: Vector search returns a confident but wrong answer (e.g. 'week 6 had the highest sales' when week 14 was actually higher) **Why it happens:** The agent only saw the few chunks that matched the query, not the full dataset **Fix:** For questions that need full context (totals, summaries, comparisons), store the data as a plain markdown file the agent reads in full instead of chunking it into a vector DB ## Level 1 CLAUDE.md router (structure shown in the video) Template shown in the video. Adapt to your project — do not copy verbatim without reviewing each line. ``` # CLAUDE.md # This file loads automatically every time you open Claude Code in this folder. # It tells the AI who you are, how you work, and where things live. ## Role [who you are / how you work — your system prompt for this project] ## Where things live (routing rules) - /context/about-me.md -> always-true background about you and how you work. Read this first. - /context/stack.md -> tools and stack you use - /context/decisions.md -> decision log (append new decisions + dates) - /projects/ -> one file or folder per ongoing project / client ## Rules - When you make a big decision, append it (with the date) to /context/decisions.md ``` ## Gotchas & Caveats - Higher level is not better. Nate runs his entire main project at Level 2 — only climb a level when you feel a real retrieval pain. - Anything you process through Claude goes to Anthropic. For private client data, use local / open-source models instead of putting everything into a cloud second brain. - Vector databases are not magic — they retrieve similar chunks, not whole documents. Don't rely on them for summaries or aggregate questions. - There is no single 'correct' folder structure. What matters is that the routing makes sense to both you and your agent. ## Key Takeaways - A second brain is just markdown files and folders organized so both you and your agent can find things again — it's harness-agnostic (works with Claude Code, Codex, Hermes). - Work backwards from the question: how you'll retrieve the data determines how you should store it in the first place. - The five levels: CLAUDE.md router + folders → LLM wiki + automemory → semantic/vector search → knowledge graphs → full autonomy. - Mix storage styles inside one second brain — markdown for full-context items, vectors for needle-in-haystack lookups, graphs for relational data. - Before blaming the AI for bad retrieval, check whether you actually got everything out of your head and into the system. ## More OpenClaw & Claude Code news [▶ Who Owns Your AI Agent? The Maintenance Skill Teams Skip in 2026 2026-06-21](https://openclawdatabase.com/news/videos/2026-06-21-who-owns-your-ai-agent/) [▶ Open Skills: Portable, Composable Agent Procedures Across Every Tool 2026-06-21](https://openclawdatabase.com/news/videos/2026-06-21-open-skills-portable-agent-procedures/) [▶ How to Set Up GLM 5.2 in Claude Code (~5x Cheaper Than Opus) 2026-06-21](https://openclawdatabase.com/news/videos/2026-06-21-glm-5-2-claude-code-setup/) [▶ Idea to Deployed AI App with Claude Code, the Vercel AI SDK, and design.md 2026-06-20](https://openclawdatabase.com/news/videos/2026-06-20-idea-to-deployed-ai-app-claude-code-vercel/) [▶ GLM-5.2 vs Opus 4.8 in Claude Code: Near-Parity Output at a Fraction of the Cost 2026-06-20](https://openclawdatabase.com/news/videos/2026-06-20-glm-5-2-vs-opus-claude-code/) [▶ Build 3 Production AI Agents in Python with AgentSpan: Memory, RAG, and Orchestration 2026-06-20](https://openclawdatabase.com/news/videos/2026-06-20-build-production-ai-agents-python/) [See all OpenClaw news →](https://openclawdatabase.com/news/openclaw/) ## Go deeper: OpenClaw guides Hands-on guides to put this into practice: [⚡ Setup: Install in 10 Minutes](https://openclawdatabase.com/openclaw/setup/) [🔐 Security Hardening](https://openclawdatabase.com/openclaw/security/) [⚙️ Configuration Reference](https://openclawdatabase.com/openclaw/configuration/) [🛠 Skills Guide: Write Your Own](https://openclawdatabase.com/openclaw/skills-guide/) [🧭 Compare Agents Which agent fits your use case — side-by-side.](https://openclawdatabase.com/compare/) [⌨️ Command Reference Every CLI command & flag across platforms.](https://openclawdatabase.com/commands/)