Transfer memory across Coding Agents

Graphify: Step-by-Step Setup Guide - Google Docs

Base on Github Repo

Descriptive Video on the Topic

Nate Herk - Andrej Karpathy Just 10x’d Everyone’s Claude Code

Proper Notes

Building an LAG Wiki in Obsidian with Claude Code

Structured knowledge document derived from the provided YouTube transcript.

Overview

This document explains a lightweight approach to building a personal knowledge base or second brain using:

• Claude Code for ingestion and organization
• Obsidian as the Markdown-based front end
• A simple LLM wiki architecture inspired by Andrej Karpathy's idea of LLM knowledge bases

The core concept is straightforward:

1. Put raw source material into a raw/ folder.
2. Ask the LLM to ingest and organize it.
3. Let it generate structured Markdown pages inside a wiki/ folder.
4. Use links, indexes, logs, and summaries to navigate and query knowledge efficiently.

This avoids the complexity of traditional RAG infrastructure for small to medium-sized knowledge systems.

Core Idea

Instead of building a retrieval system around embeddings and vector databases, the workflow uses a folder of Markdown files as the knowledge substrate.

Key characteristics

• Knowledge is stored in plain Markdown files
• Relationships are represented through links and backlinks
• The LLM maintains:
  • index pages
  • summary pages
  • logs
  • entity and concept pages
• The system compounds over time as more material is ingested

The result is a knowledge base that behaves less like a chat history and more like a persistent, evolving wiki.

What the System Looks Like

The transcript describes two example knowledge systems:

1. YouTube Knowledge System

A vault containing transcripts from recent YouTube videos, automatically turned into:

• source pages
• concept pages
• tool pages
• framework pages
• backlink-rich notes

Each video can include:

• video title
• tags
• source link
• raw transcript file
• explanation of the video
• key takeaways
• links to related tools, frameworks, and ideas

2. Personal Second Brain

A separate vault used for personal and business context, such as:

• business initiatives
• employee context
• quarterly priorities
• meeting notes
• internal summaries
• project knowledge

This second system can then be connected to an AI assistant so the assistant can selectively read from the wiki when needed.

Why This Matters

Traditional AI chats are ephemeral. The knowledge generated during a conversation is usually lost unless manually preserved.

This approach changes that by making knowledge:

• persistent
• structured
• queryable
• cumulative

The transcript frames this as knowledge that compounds like interest, because every new source improves the value of the system instead of disappearing after one interaction.

What he did

• Create a new Obsidian vault
• Open this value in VS Code
• Open terminal
• Launch Claude Code
• Go to Karpahty's LLM Wiki
• Copy LLM Wiki contents
• Paste in Claude code prompt with its own prompt

• After entering prompt Claude will create raw and wiki folders structure
• Then he went on to and copy a research page and put in obsidian or use the extn. web clipper and put to raw/ folder
• Explain to claude code what you just did (mention about the page) and ask to ingest it.
• After wiki the claude or the llm will break it into many .mds files and basically creates the knowledge source out of it. - chunking
• Explain to LLM answer to question it asks and depending upon your answer it will do chunking