I started vibe coding about a year ago. AI would build things that worked, but when something broke I had no idea where to look. I realized I was shipping code I couldn't explain. The turning point was watching Karpathy talk about his autoresearch loop: experiment, measure with a scalar metric, keep what works, discard what doesn't, compound. I thought, that's exactly what's missing from my coding workflow. There's no memory. No compounding. Every AI session starts from scratch. So I built a simple system: an Obsidian vault where my AI reads rules before every task and writes a walkthrough after. When something breaks, the walkthrough proposes a rule. I promote it. Next session, the AI follows it. The rules folder grows. The mistakes stop repeating. The part that surprised me most was the NotebookLM integration. I export my walkthroughs weekly and upload them. NotebookLM generates audio overviews of my own code. I listen while walking. Patterns started clicking in ways that reading never achieved. I built the entire product (vault, landing page, distribution) in one session with Claude Code. The vault itself is open source on GitHub. The Pro version adds example walkthroughs, an eval harness, and instruction files for Cursor/Copilot/Windsurf so the loop works with any AI tool. Would love to hear if anyone else has tried building persistent memory systems around their AI tools.

CodeGlass — Evolve Together

Your AI remembers every mistake and never makes it twice.

Your AI remembers every mistake and never makes it twice.