From Second Brain To LLM Wiki: How I Built a Compounding Knowledge Base with AI

A few years ago I wrote about building a second brain — offloading information from my head into Logseq (which uses markdown, so it’s compatible with Obsidian too), using Tiago Forte’s P.A.R.A. method to tame the chaos. That system changed how I work. But with recent advances in AI and trying to stay on top of emerging trends I needed help.

I was drowning in AI sources. Podcasts, docs, GitHub repos, blog posts — all about the agentic AI ecosystem. I’d read something brilliant, file it in Obsidian, and three weeks later forget it existed. Worse, I couldn’t see the connections between sources. How does Kiro’s context management relate to Anthropic’s skill standard? Where does Fabric fit in the stack? My second brain had the raw material, but no synthesis.

Then I found Andrej Karpathy’s LLM Wiki pattern — and everything clicked.

The Problem with RAG (and My Notes)

The core insight is deceptively simple: Wiki > RAG.

RAG (Retrieval-Augmented Generation) rediscovers knowledge from scratch on every query. You ask a question, it searches your documents, pulls relevant chunks, and synthesizes an answer. It works, but nothing compounds. Ask the same question tomorrow and it does the same work again. There’s no accumulation.

My Obsidian vault had the same problem at a different scale. The notes were there, but the synthesis lived only in my head. I was the one maintaining cross-references, spotting contradictions, connecting ideas across sources. And I’m not great at that when I have 17 sources across tools, standards, methodologies, and practitioner insights.

Karpathy’s idea: let the LLM do the maintenance. Humans curate sources, direct analysis, and ask good questions. The LLM does everything else — summarizing, cross-referencing, filing, and bookkeeping.

How It Works: Three Layers, Three Operations

The architecture is clean:

Raw Sources (immutable)  →  Wiki (LLM-maintained)  →  Schema (CLAUDE.md)
   you add these              LLM writes these         you + LLM co-evolve

1. Raw sources — Immutable documents. The LLM reads them but never modifies them.
2. The wiki — LLM-generated markdown. Entities, concepts, comparisons, synthesis. The LLM owns this entirely.
3. The schema — Instructions telling the LLM how the wiki works. Structure, conventions, workflows.

And three operations:

• Ingest: Drop a source in, the LLM processes it, creates or updates 10-15 wiki pages, updates the index, appends to the log.
• Query: Ask a question, the LLM reads the index, finds relevant pages, synthesizes an answer — and optionally files the answer back as a new analysis page.
• Lint: Health-check for contradictions, orphan pages, stale claims, missing cross-references.

What I Built

I applied this pattern to map the agentic AI landscape. The result is my LLM Wiki — 17 sources ingested, producing 18 concept pages, 16 entity pages, and 4 synthesized analyses.

The sources span the full stack:

• Tools: Scion, Kiro, Claude Code, Fabric, Paperclip, Promptfoo, NotebookLM
• Methodologies: Spec-Kit, BMAD Method, Ten Pillars of Agentic Skill Design
• Standards: Agent Skills Standard, MCP Protocol
• Evaluation: Anthropic’s eval guide, Promptfoo
• Practitioner insights: Real-world patterns from podcasts and skill pipelines

What emerged was something I couldn’t have built manually — a five-layer stack that kept surfacing across all the sources:

Layer	Tools	Focus
Company	Paperclip	Org charts, budgets, governance
Methodology	Spec-Kit, BMAD	Specs, plans, quality gates
Infrastructure	Scion	Containers, runtimes, isolation
Tool	Claude Code, Kiro	Agentic loop, skills, MCP
Pattern	Fabric, Agent Skills Standard	Curated prompts, composable skills

I didn’t design this taxonomy. It emerged from the cross-referencing the LLM did during ingest. That’s the compounding effect in action.

The Obsidian Connection

Here’s what surprised me: the Obsidian workflow I’d built for my second brain was the perfect human layer for this pattern.

Obsidian handles what I’m good at — capturing sources as I encounter them, journaling my thinking, tagging things by project and area (P.A.R.A. still works). The LLM Wiki handles what I’m bad at — maintaining cross-references across 50+ pages, spotting when a new source contradicts an old one, synthesizing themes across everything I’ve read.

The workflow looks like this:

1. I discover a source (podcast, repo, blog post) and capture it in Obsidian
2. When I have enough context, I feed it to the LLM Wiki for ingest
3. The LLM creates/updates entity pages, concept pages, and cross-references
4. I query the wiki when I need synthesis (“How do these tools compare on context management?”)
5. Good answers get filed back as analysis pages — they compound too

The human curates. The LLM maintains. Neither could do the other’s job well.

Why Wikis Usually Fail (and Why This Doesn’t)

Vannevar Bush imagined the Memex in 1945 — a personal knowledge store with associative trails between documents. His vision was closer to this than to what the web became: private, actively curated, connections as valuable as documents.

The part he couldn’t solve was who does the maintenance.

Humans abandon wikis because the maintenance burden grows faster than the value. You start enthusiastic, cross-reference everything, then life happens and the wiki rots. I’ve done this with Confluence, Notion, and yes, even Obsidian vaults that got too big.

LLMs don’t get bored. They don’t forget cross-references. They can touch 15 files in one pass. The maintenance cost is near zero.

Eight Themes That Keep Surfacing

Across all 17 sources, the wiki surfaced eight recurring themes (the full analysis is in the cross-source themes page):

1. Context beats clever prompting — Progressive disclosure, context documents, selective loading. This was the strongest consensus across all sources.
2. Composition over monoliths — Every tool chose small, focused, composable units.
3. The human stays in the loop — but how much? — A spectrum from “always interactive” to “days of autonomy.”
4. Five orchestration layers emerging — The stack I described above.
5. Two methodology philosophies — “Specs before code” vs. “Expert collaboration over autopilot.”
6. Memory is the unsolved frontier — Persistent context compounds value and errors.
7. Open standards winning — MCP + Agent Skills Standard as a two-layer open substrate.
8. Evaluation is the weakest link — Everyone agrees it matters; nobody has a complete answer yet.

I couldn’t have synthesized these manually. Not across 17 sources. The wiki did it because every ingest pass updates every relevant page, and the connections accumulate.

What I Learned

Building this wiki taught me a few things that go beyond the technical:

Start with sources you care about. The wiki is only as good as what you feed it. I chose sources I’d already read and had opinions about — that made the curation meaningful.

The schema is everything. The CLAUDE.md file that defines how the wiki works is the most important artifact. It’s effectively a skill — instructions that shape agent behavior for a specific domain. Getting this right took iteration.

Query your own wiki. The best analyses came from asking the wiki questions I genuinely wanted answered. “How should I evaluate an agent skill?” became a full analysis page with a practical framework.

Lint regularly. The lint operation catches contradictions and gaps. It’s like code review for knowledge.

From Second Brain to Compound Brain

When I wrote about building a second brain, the goal was simple: stop losing information. Offload it, organize it, retrieve it when needed.

The LLM Wiki takes that further. It’s not just storage and retrieval — it’s synthesis and accumulation. Every source makes every other source more valuable because the cross-references compound.

If you’re already using Obsidian or any PKM tool, you have the human layer. The LLM Wiki pattern gives you the machine layer. Together, they’re closer to Bush’s Memex than anything I’ve used before.

The wiki is live: blog.imfsoftware.com/llm-wiki/docs

Browse it. See how 17 sources become a connected knowledge base. And if you build your own, I’d love to hear what emerges.

---

Resources

• LLM Wiki — Agentic AI Landscape
• LLM Wiki Pattern (Concept Page)
• Andrej Karpathy’s Original LLM Wiki Idea
• Cross-Source Themes Analysis
• Building a Second Brain (my earlier post)
• Obsidian

image-ref: https://unsplash.com/photos/gold-and-purple-beaded-necklace-QJuzFcO-p3A tags:

• ai
• knowledge-management
• obsidian
• llm