Andrej Karpathy just endorsed one of the most interesting AI experiments of the year: Farzapedia --- a personal Wikipedia built from 2,500 diary entries, Apple Notes, and iMessage conversations, auto-organized into 400 detailed wiki articles.

But Karpathy’s excitement wasn’t about the novelty. It was about the architecture. He called it a superior approach to AI personalization, and his reasoning matters for anyone thinking about how to build an effective AI workflow.

The Problem with “AI Memory”

Most AI products today promise that your AI assistant “gets better the more you use it.” Your conversations are analyzed, patterns are detected, and the AI silently builds an understanding of who you are.

The problem? It’s a black box.

You can’t see what the AI thinks it knows about you. You can’t correct wrong assumptions. You can’t share your knowledge context with a different AI tool. And when you switch providers, you start from zero.

Karpathy identified three properties that make a personal wiki fundamentally better:

Explicit: You can see exactly what the AI knows or doesn’t know about you. The knowledge is viewable, navigable, and editable --- not buried inside an opaque model.
Portable: The data is yours. It’s just files. You can move it between tools, back it up, or share parts of it with others.
Composable: Different AI tools can read the same knowledge base. Your research doesn’t belong to ChatGPT or Claude --- it belongs to you.

From Diary to Wikipedia --- and From Web to Wikipedia

Farzapedia started from personal notes. But most people’s knowledge doesn’t live in their diaries --- it lives in the web pages they read every day.

Think about it: articles you’ve researched, product pages you’ve compared, documentation you’ve referenced, tutorials you’ve followed. That’s your working knowledge, scattered across hundreds of browser tabs and bookmarks that will eventually be forgotten.

What if you could turn all of that into your own personal Wikipedia?

Building Your Wiki LLM in Practice

The approach is straightforward:

1. Capture Everything Worth Remembering

Instead of bookmarking pages (which you’ll never revisit), convert them to clean, structured Markdown as you browse. This preserves the actual content --- not just a URL that might break or change.

With Save, every webpage becomes a Markdown file with one click. The AI extracts the meaningful content, strips the noise, and structures it for long-term use.

2. Organize Into Knowledge Bases

Just like a Wikipedia has categories, your personal wiki needs structure. Group your saved pages into knowledge bases by topic:

Save Vault/
  React Research/
  Competitor Analysis/
  Product Design/
  Industry Trends/
  Cooking/

Each knowledge base becomes a category in your personal encyclopedia.

3. Make It Searchable by AI

Here’s where it gets powerful. When your personal wiki is structured as Markdown files in a local folder, AI assistants can search and reference it directly.

With Save Vault and its built-in MCP server, Claude can:

Search across all your saved pages
Read any article in full
Cross-reference information between different knowledge bases
Answer questions grounded in your curated knowledge, not generic training data

When you ask Claude a question, it checks your personal wiki first. If you saved an article comparing database architectures last week, Claude will reference that specific article rather than giving you a generic answer.

4. Let It Compound

The real magic is in the compounding. Every page you save makes your personal wiki more complete. Over weeks and months, you build a knowledge base that no AI could replicate from scratch --- because it reflects your specific interests, research history, and professional context.

This is what Karpathy means by “explicit personalization.” It’s not an AI guessing what you care about from conversation patterns. It’s a structured, visible, editable body of knowledge that you built intentionally.

Why Markdown Is the Right Format

Karpathy’s emphasis on data ownership isn’t accidental. The format matters.

Markdown is universal: Every AI tool can read it. It’s not locked into any platform.
Markdown is human-readable: You can open any file in any text editor and read it.
Markdown is version-controllable: You can track changes over time with Git.
Markdown is lightweight: Thousands of articles take up almost no disk space.

This is why tools like Obsidian, Claude Code, and Save all speak Markdown natively. It’s the lingua franca of the AI-native knowledge stack.

The Personal Wikipedia Workflow

Here’s the complete workflow in practice:

Browse the web normally. When you find something worth keeping, click Save.
Save converts it to Markdown --- clean, structured, with metadata preserved.
Save Vault stores it locally in your knowledge base folders.
Claude reads your vault via MCP when you ask questions, grounding answers in your saved knowledge.
Your wiki grows automatically with every page you save.

No setup. No database. No subscription to a knowledge management platform. Just Markdown files on your computer, accessible to any AI tool you choose.

From Consumer to Creator of Knowledge

Karpathy’s Wiki LLM concept represents a shift in how we think about AI personalization. Instead of being passive consumers of AI --- feeding our data into opaque systems and hoping they learn --- we become active curators of our own knowledge.

Your personal Wikipedia is explicit, portable, and owned by you. And it makes every AI interaction better because the AI is working with your knowledge, not just its training data.

The smartest AI strategy for 2026 isn’t using a smarter model. It’s building a smarter knowledge base.

Save converts any webpage to clean Markdown and stores it in your local knowledge base. Combined with Save Vault’s MCP server, Claude can search and reference your saved knowledge directly. Try Save for free.