The numbers
Embed coverage
Search works two ways at once: keyword match, and semantic match against an embedding of each page. Embeddings are generated automatically when a page is saved, and an hourly maintenance pass picks up any that were missed. Coverage is the share of pages (of those with anything to embed) that semantic search can currently see. A healthy brain reads 100% and stays there - green from 90%, amber from 60%, red below. A brief dip after importing a lot of pages at once is normal; the hourly pass catches it up. A low number that does not climb on its own is worth fixing: press Run dream to run the maintenance pass now, and if the number still will not move, contact support.Oversized pages
A page is one topic, and the size rule is 300 to 1,500 words. Health counts the pages over 2,000 words and lists them, because a page that long answers questions badly: search finds it, but the answer is buried. A page that grew by appends is folded for you within the hour: its dated lines older than six months move, unchanged, to a history page beside it, linked both ways, and so do the oldest of the rest while the page is still over the line. The page keeps its current state and at least its twenty newest lines. A page still on the list after that has twenty long lines, and the count is information rather than a fault. A reference page that is really several topics is best split by hand with an overview left at the original address. Nothing rewrites a page for you: the model rewrite Cortex queued until 7 October 2026 was withdrawn because it lost detail on exactly the pages that grow. Folder index pages, history pages and your task page are never counted.Is it still being fed?
The numbers above say whether what went in is findable. They do not say whether anything is still going in. A brain set up once can quietly stop receiving: the session hook was lost with a rebuilt laptop, the meeting tool’s webhook was switched off, a credential expired, or items have sat in the inbox for a fortnight. Ask your agent: “Run the health check on my brain.” It reads one report and answers, worst first, with the one thing that fixes each:
Those are the defaults. If your brain expects more or less, say so on its instructions
page under a heading
## Expectations, in plain sentences: “Session transcripts arrive most
days (stale after 2 days)”, “Granola transcripts arrive on weekdays through the Granola
source”, “No meeting tool is connected to this brain”. Your agent reads that section first
and grades against it, so a feed you do not use is never reported as missing. The check
never changes anything: it tells you what to renew, switch on or clear, and you decide.
The same report is there without an agent: cortex health --brain <slug> from the
command line prints it and exits with a
code a scheduled job can alert on, and the API
returns it as JSON.
The buttons
- Run dream runs the hourly maintenance pass immediately: resolves links, fills in any missing search data and embeddings, and refreshes the brain’s front page if you have not edited it. Harmless to press whenever - repeated presses coalesce into one pass.
- Repair structure rebuilds what the brain’s template provisioned: its instructions, schema and folder index pages. Use it if the Health page warns about stuck page writes or template folders are missing. It never touches pages you or your agents wrote.
Try it
- Open a brain’s Health page and check embed coverage is green.
- Fix one broken link: open the listed page, correct the
[[wikilink]], watch the count drop. - Press Run dream and refresh after a minute to see the counts settle.
- Ask your agent to run the health check and read which feeds it found.