What is worth remembering

Hivemind is deliberately conservative about what it captures on your behalf. The things that pay off most are the ones you state on purpose.

Worth storing

Decisions and their reasoning. The reasoning is the part that gets lost, and the part that stops the next thread redoing work you already ruled out.

We kept pgbouncer rather than raising the pool size, because a bigger pool just masks the retry storm.

Constraints. Rules that should shape every answer.

Never deploy on a Friday afternoon. Production caps at three deploys a month.

Preferences. How you want things done, stated once.

Tabs, not sidebars. British spelling in user-facing copy.

Gotchas. Anything that cost you an hour once.

ops/deploy.sh refuses to run with a dirty working tree, and the error does not say so.

Not worth storing

• Anything in the repository. File contents and structure are better read than remembered, and go stale the moment someone commits.
• Transient state. "The build is running" is worthless in an hour.
• Long pastes. Several short, self-contained statements retrieve far better than one wall of text, because each can be matched on its own.

Writing them well

Make each one stand alone. It will surface without whatever you were saying around it.

The staging database is Postgres 16, in eu-west-1.

It's on 16 now.

Include the words you would search by. If you would ask "which database does billing use", the memory should contain both database and billing.

State the current fact, not the change. Write "deploys are capped at five a month" rather than "we raised the cap", the first is still true in six months and supersedes the old value cleanly.