Skip to content

Best practices

  • One source at a time. Connect, small ingest, dream --dry-run, dream, read the brief. Then the next source. Five at once is undebuggable.
  • Dry-run before every dream that matters. It prices the pass and shows the bundles the model will see. Surprises belong in preview, not in bills.
  • Read the gate before widening. memcal gatecheck tells you whether a new source's traffic is signal or noise shaped like signal. Tune senders (memcal senders), top-tier people (memcal top), and mutes before paying for a full pass.
  • Mute for noise, block for never. Muting keeps archive rows and stops model spend; blocking drops the sender entirely. Platform mutes are recorded, not fought.
  • Correct in the open, in the day. A wrong row fixed with a typed write now beats a perfect dream tonight — and the correction survives the night. Never let a wrong answer sit because "dream will fix it."
  • Name people early. memcal who --adopt is free; an afternoon of --confirm and --split beats a month of misattributed rows. Contacts-backed identities outrank all future guesses.
  • Keep the wiki lazy. Pages earn existence from evidence, not from contact imports. Facts on pages, obligations in to-dos, uncertainty in questions — each in its typed home.
  • Schedule the night job. memcal schedule install makes freshness and dream automatic: catch-up collection on wake, the full pass when owed, cheap checks otherwise.
  • Review weekly. memcal review plus the Runs tab: what it wrote, what it asks, what it could not resolve. Ten minutes here prevents zombie to-dos and stale pages.
  • Measure plumbing free, quality deliberately. The integration benchmark runs in a second on every change; the model layer and the seven-day experiment spend real money and run on declaration, frozen corpora, and held-out lives.