A self-tending knowledge base
A knowledge base that tends itself.¶
Sources in. Cited pages out. The agent reads everything, writes it into a cross-linked web of citable pages, and keeps that web consistent as new sources arrive.
Most personal knowledge bases decay. Notes pile up, links rot, and the insights you captured last year become ghosts — present in the file system, invisible in practice. Corpus is built on a different premise: the LLM agent is the librarian. It reads every source, extracts what matters, and writes it into a self-organizing web of cross-linked, citable pages — then keeps that web consistent as new sources arrive.
The result compounds. Each new source enriches the pages that are already there. Each query can surface what the corpus knows and identify what it doesn't — automatically queuing the gap for the next ingest run.
The pipeline · collect → ingest → cite
Every arrow is automated. Every corpus page cites its sources. Every source is stamped once it's processed — so nothing gets ingested twice.
The corpus, mapped¶
Each dot is a page; the larger nodes are the eight domain hubs, and the lines are the citations and cross-links between pages. This map is generated from the live corpus on every commit — it grows as the corpus does.
Drag to explore · hover a node for its title · 416 pages · 8 domains · 1740 sources.
01
Why a self-tending knowledge base beats a folder of notes — and why provenance is the one non-negotiable.
02
The end-to-end pipeline: collect → inbox → cluster → ingest → verify. How sources become cited pages.
03
Five intake channels — email, YouTube, PDF, Obsidian vault, and web — each with its own harvesting logic.
04
The autonomous agent runtime: ingest, lint, adapt, and eventually dream. How the corpus tends itself overnight.
05
Schema, frontmatter spec, domain rules, the op log, and the operating manual (CLAUDE.md) that governs it all.
Wire the collectors to your own accounts, run your first ingest, and install the nightly schedule.
Why this exists¶
Information that isn't cited can't be audited. Information that isn't cross-linked can't be found. And information that lives only in a chat history vanishes the moment the session ends.
Corpus treats every source as raw material. The agent's job is not to answer questions — it's to build a durable, searchable, linkable layer of derived knowledge that can answer questions, today and years from now. It's a compounding asset, not a conversation log.
Design principle
"Without provenance, the corpus becomes lossy compression you can't audit."
The whole system is governed by a single co-evolved operating manual that specifies path isolation, page types, ingest steps, lint checks, and anti-drift rules. The agent runs autonomously within those rules; the maintainer co-evolves the rules over time.
Next: The Idea — the philosophy behind a self-organizing knowledge base.