Why We Picked 'LLM Knowledge Base' Over 'Memory Layer' — A Category Anchor Decision

On April 3, 2026, Andrej Karpathy posted a public gist describing how he uses a personal "wiki / knowledge base" that he manually copy-pastes into every LLM session. Within 48 hours, "LLM Knowledge Base" and "LLM Wiki" were trending headlines on VentureBeat and AnalyticsIndiaMag.

This is the kind of category-formation moment that happens once every two years. We had a landing page hero pointing the wrong way, and we had a window of maybe two weeks to ride the wave before someone else anchored it.

This post is the decision log of how we picked our new category word — and why we rejected the one that, on paper, looked obviously better.

The starting point: "AI-Native Second Brain" is dead

KnowMine's old hero said:

Your AI-Native Second Brain.

It was fine in 2024. By 2026 it's a graveyard. "Second brain" is a Tiago Forte keyword owned by Notion, Obsidian, Mem, Reflect, and a dozen others. The note-app red ocean. Every "AI-native note app" sounds the same. We needed out.

We brainstormed three replacements:

• A. "Your knowledge, finally portable." — too abstract, no category anchor
• B. "The AI that doesn't forget you." — negative framing, anchors on the problem not the solution
• C. Bilingual hero with Chinese on the same screen — Google would penalize the EN page for low keyword density

All three failed. Time for option D.

Round 2: "Memory Layer" — the obvious answer

The most-talked-about new category in AI infra during 2026 has been the memory layer for AI agents. Mem0, Zep, Letta, Supermemory, LangMem — every one of them frames their product the same way. Mem0 just raised a $24M Series A on this exact wording.

So obviously we should grab "Memory Layer," right? The user even pushed back when I first proposed something else: the "second brain" framing has普通人 value, don't throw away the emotional resonance.

I steelmanned it and almost shipped it. Then I made one phone call to the data and it fell apart.

Round 3: the evidence that killed Memory Layer

Three things, in order:

1. Karpathy himself avoids the word "memory."

I went back and read his original gist with a highlighter. He uses wiki, knowledge base, persistent compounding artifact. He uses the word "memory" zero times in the body. The ecosystem coverage that followed (VentureBeat, AnalyticsIndiaMag, Hacker News thread) all picked up "LLM Knowledge Base" or "LLM Wiki." Not one major outlet called it memory.

The category-word giver of this wave explicitly didn't pick "memory."

2. Mem0 owns "Memory Layer for AI Agents" — fully.

Their official one-liner is literally "Universal memory layer for AI Agents." They have $24M in fresh capital to defend that wording on Google, on conference stages, in every podcast. Trying to wedge a C-end personal product into that B2B infra slot is fighting a battle we will lose, with users we don't even want.

3. Karpathy and Mem0 live in parallel universes.

I audited four weeks of coverage. No journalist, no Twitter thread, no podcast has put "Karpathy's wiki idea" and "Mem0's memory layer" in the same sentence. They're serving different audiences (individual builders vs. agent infra teams), with different solutions (a markdown file you copy-paste vs. an API your agent calls), and the discourse never crosses over.

If we picked "Memory Layer," we'd land in mem0's universe — and lose. If we picked something Karpathy-adjacent, we'd land in his universe with no incumbent.

The pick: "LLM Knowledge Base"

Five reasons it won:

1. Karpathy's exact word. Maximum SEO tailwind on a phrase that just got blessed by the most-cited voice in the space.
2. No commercial incumbent. sage-wiki and CRATE are open-source projects, not category-owning products. The slot is empty.
3. Our moat fits perfectly. Karpathy's solution is manual copy-paste of a markdown file. CRATE and sage-wiki are local-only markdown stores. KnowMine is the first MCP-native LLM Knowledge Base — Claude, ChatGPT, Cursor, and any MCP client can read and write to it mid-conversation. That's a genuine product gap, not marketing spin.
4. We dodge the $24M battle. Different users (C-end individuals, not agent infra teams), different scenarios (knowledge work, not API calls), different stack. Mem0 and us don't even compete.
5. The scenario split holds up. Traditional notes start with "open the app and write." KnowMine starts with "you're already in an AI conversation." Two structurally different markets — and the new framing names ours precisely.

What changed on the page

• Hero EN: The LLM Knowledge Base your AI can read and write.
• Hero ZH: 你的 AI 能直接读写的知识库。
• Subtitle: "Not another note app. A protocol-native knowledge base where Claude, ChatGPT, Cursor, and any MCP client can save, search, and grow your long-term knowledge — in the middle of a conversation."
• New second screen: a logo wall showing every MCP client KnowMine works with natively.

The lesson

When a new category word breaks through, the default move is to grab the loudest-sounding adjacent word. Don't. Read the actual source carefully — what word did the wave's originator use? — and look at who already owns the alternatives. The word with no incumbent and the strongest-cited backer wins.

We almost shipped "Memory Layer" because it felt strong. The data killed it in 90 minutes. That's a cheap save for a positioning decision the entire product is going to wear for a year.

If you're building an AI knowledge tool, the category word matters more than the feature list. Pick the one that's empty, and the one the loudest voice in the room just used. Sometimes those are the same word.

---

KnowMine is the first MCP-native LLM Knowledge Base — Claude, ChatGPT, Cursor, and any MCP client can save, search, and grow your long-term knowledge mid-conversation. Try it free →