I read 1,903 things.
I remember all of them.

The problem was never finding things. It was keeping them.

A modern information diet is a firehose: 187 newsletter senders, plus a constant stream of papers, repos, YouTube transcripts, and WeChat threads. You read it, you nod, it scrolls past — and a week later it's gone.

• Save the tenth article on an idea and it never makes the first nine smarter — they sit in separate tabs, separately forgotten. A bookmark is storage that pretends to be memory.
• The reframe that started everything: stop treating reading as the product. The product is the structured residue reading leaves behind.
• So the goal isn't a better read-it-later app. It's a compiler — something that turns each source into a durable, linked node and files it next to everything related.

An unread bookmark is a debt. A compiled note is an asset.

So I built a compiler for the things I read. Raw in. Wiki out. Outputs filed back.

The whole architecture fits in one line: [Raw Sources] → [LLM Compiler] → [Wiki] → [Obsidian Viewer], with Q&A outputs promoted back into the wiki whenever they're reusable.

• Claude Code is the compiler; a local Obsidian markdown vault is the substrate. Everything is plain markdown on local disk — git is the audit log, 286 commits so far.
• The compiler treats untrusted content as data, never instructions — it extracts factual claims and citations; it never executes what it reads.
• The crucial inversion: the LLM owns the wiki and writes it; the human reads it and rarely edits. The folder is the memory; the model is what keeps rewriting it.

Memory needs a shape. This one has five.

Information doesn't compound when it's a flat pile. It compounds when raw sources, compiled knowledge, answers, projects, and prompts each live in their own layer — and link to each other.

Seven doors in. Eight workers per wave. One language out.

Each source type gets a dedicated pipeline — a Claude Code "skill" that knows how to extract, translate, dedup, and file that kind of content. Whatever the format, whatever the language, what lands in the wiki is clean English markdown.

• Newsletters arrive by the hundred, so they get a wave architecture: waves of ≤8 parallel workers map-extract, then one reducer per wave compiles — never one giant reducer holding everything.
• Non-English in, English out, always. A WeChat essay arrives in Chinese; the raw note keeps the original verbatim with a translation appended — the wiki only ever speaks English.
• A three-key dedup (Message-ID · canonical URL · thread ID) survives multi-account forwarding and resends. 187 senders feed in; the top feed is Paul Krugman at 159 notes.

---
type: source
source_type: newsletter
title: "Six Lessons on Founder Mode"
author: "..."
date: 2026-06-17
tags: [startups, leadership]
projects:
  - "[[Founder Mode Research]]"
---
# extracted, deduped, linked

Knowing it's stored is worthless if you can't find it.

Storage is the easy half. The hard half is recall — pulling the right four pages out of 620 when you ask a fuzzy question. So retrieval fuses three different ways of being right.

1. Hybrid by design. BM25 catches exact terms, dense embeddings catch meaning, backlink expansion catches the neighborhood — Reciprocal Rank Fusion blends all three into one ranking.
2. On-device, disposable index. Markdown stays the source of truth; the vector index is a rebuildable cache that never leaves the laptop. Quality is tracked as hit-rate@k against a hand-built golden-query set — "search got better" becomes a claim with evidence.

The vault keeps a heartbeat — and I'm not the one beating it.

The best knowledge system is the one you don't have to remember to use. Three scheduled passes run on a local cron, against a local Gmail bridge no cloud worker can reach. That isolation is the point.

620 pages across nine domains, with the busiest concepts pulling the most threads.

The wiki's shape isn't designed — it emerges from where attention actually went. The domain mix and the gravity wells are a portrait of one engineer's curiosity, drawn by the links themselves.

Attention has a shape, and the links draw it. Nothing here was placed by hand — the topology is just the residue of where the reading went.

Ten weeks of compounding, on one card.