"Why did we do multi-tenancy row-level?" — the engineer who decided is gone, and the meeting notes still hold his exact sentence and its revision trail

What's piled up

Five years of it, all in markdown, all searchable. What search hands back is a list of lines containing the string "multi-tenancy." Which of those lines was a decision and which was an opinion isn't written anywhere in the list.

• 200-odd weekly engineering meeting notes — one a week since March 2021. Most state the date on the first line.
• 30-odd architecture decision records — under docs/adr/. Written as context, options, decision, consequences, until the format started drifting around 2023.
• 60-odd Slack excerpt notes — whole threads copied and pasted into a note. About 30 of them never say what date the conversation happened.
• 4,000-odd markdown notes across the workspace — beyond those three kinds, there's onboarding docs, retros, and spec drafts mixed in.
• Sentences from three people who left — the words of the people who actually made the calls are still sitting in the body of the meeting notes.

How it works today

The CTO's current procedure goes like this. Search Slack for "multi-tenancy," skim the 400-odd results from the top, and look for messages that read like decisions. Then open docs/adr/ and click into whatever looks relevant. There's one called adr-011-tenancy.md, written in 2021, and whether that decision changed afterward is not written anywhere inside it.

Two places break. First, in the search results, "I think splitting schemas would be better" and "Decision: we're going row-level" come back weighted the same. Second, there's an exception buried in a June 2023 meeting note — three enterprise accounts get carved out onto separate schemas — and the 2021 ADR shows no trace of it. So the new engineer reads a dead decision as if it were current.

I got the same question 18 months ago. And now I can't remember where I wrote down the answer it took me three days to find.

How you set it up

Nothing new gets written. You put in the 290-odd markdown notes that already exist, and you teach it this team's vocabulary once.

How you ask

"Find every decision we made about multi-tenancy. Newest first."
→ 6 decisions. From 2025-03 "the three enterprise accounts move to
   separate schemas" back to 2021-11 "tenant isolation goes row-level".
   Each line carries who said it and a link to that meeting note.

"Any of those get reversed later?"
→ 2. The 2021-11 decision shows struck through with a "Superseded" badge,
   and a "Superseded by Architecture sync (2023-06)" link opens that note.
   The 2022-08 "one connection pool per tenant" decision was superseded in 2024-01.

"What does it say we gave up by going row-level?"
→ 4 claims, 2 stances. "We can't hold up if an isolation audit lands" — Jaeho,
   against badge. "Migration costs a fifth of what schema separation does" — Jihoon,
   favor badge.

Two things make that answer possible. One, when extraction reads a meeting note it holds onto the author's own sentence. The row-level tenancy entity page has an "In your words" section where decisions, questions, claims, and quotes sit grouped by kind. Jaeho's words carry the attribution that says Jaeho said them, and stances wear a favor/against badge.

Two, you can check the basis on the spot. On the entity page, hover a connection chip and the source sentence behind that fact is cited as "Architecture sync (2023-06).md · line 14". Click the note link in the answer and the original meeting note opens. What the new engineer reads isn't a summary the CTO rewrote — it's the sentence Jaeho actually typed into the 2021 note.

What changes