Knowledge Bases Need Garbage Collection
Every note you add is a note someone later has to read past. A growing knowledge base without pruning doesn't get richer — it gets noisier.
Got the voice. The style guide confirms the rules you gave me, and the sample post shows how Jon actually executes them: open on a named system + a real moment, own the wrong assumption, gloss jargon inline, land the takeaway, vary paragraph length. The original body has no <Tech> block, so nothing verbatim to carry — just facts and voice.
Here’s the rewrite:
I caught a rule in my knowledge base pointing at a service I’d shut down long ago — and it was still marked active. Still surfacing in every search my Claude Code sessions ran. Still half-trusted, because nothing had told it to stop.
A knowledge base only ever grows. You add notes, rules, docs, decisions. Nothing ever gets removed. And the assumption I’d been running on was that more was better — more captured context, more history, more coverage.
I was wrong.
Past a point, every note you add is a note that someone — or some agent — later has to read past to find the live one. Stale entries that reference dead systems don’t just sit there harmlessly; they pollute retrieval (the search an agent runs to pull the relevant notes), contradict the current truth, and quietly lower the signal of the whole collection.
So the base needs a garbage collector — the automatic routine that frees up space by reclaiming what’s no longer in use — and it needs an opinion. Mine, borrowed: default to delete. The burden of proof is on keeping a thing, not on removing it. A note with no live purpose gets archived, not kept “just in case.”
Two mechanics make it real. First, lifecycle metadata on everything — a status and a review-by date attached to each note; an entry with no review date is itself a defect. Second, a recurring sweep that flags dead-infrastructure references, merges duplicates that say the same thing, and pushes anything past its review date toward repeal. Archive with a tombstone — a marker left in place of the deleted entry, recording that it existed — rather than hoard. The audit trail survives; the active set stays lean.
The goal isn’t a smaller knowledge base.
It’s a knowledge base where everything still in it is still true.