I deleted half my Obsidian vault. The graph got better.

Pruning a knowledge graph is not the same as pruning a codebase. Here is what changed and what I missed.

What I deleted

Six months of daily notes. Every “meeting note” that never got atomized. Every “inbox” note that became a graveyard. The “someday” folder. All of it — about 600 notes.

What happened to the graph

It got denser. Counterintuitively, removing low-value nodes increased the average edge count per remaining node. The signal-to-noise ratio of the graph visualization went up.

More practically: grep started returning useful results again. Before the deletion, every search returned 40 hits. After, it returned 8 — and 7 of them were actually relevant.

What I missed

Three things I deleted turned out to be the only record of things I needed six weeks later:

1. A note about a specific nginx config quirk I’d solved
2. A conversation fragment about Terraform state locking
3. A half-finished diagram of a network topology

The nginx thing I re-solved in 20 minutes. The Terraform thing took an hour. The diagram is just gone.

The lesson

Delete with intent. Don’t delete because something looks like clutter. Delete because you’ve verified it’s not the only copy of something you might need.

The right deletion heuristic: if the note has zero outgoing refs, it’s either an atom (keep it) or a dead end (delete it).