You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Facts on a database-only source: keep the rows out of the entity page body (or bound them)
#8973
On a managed Postgres brain (gbrain 0.60.565, database-only source, write-through off), every remembered fact is added as a row to its entity page's ## Facts fence. On a hub page, the fence ends up as most of the page and most of the write cost:
measured 9 Nov 2026
company hub page
311 KB body, of which 247 KB (79%) is the Facts fence: 891 rows, about 275 bytes each
reading it
about 75k tokens for an agent that opens the whole page, most of it rows that recall already serves from the facts table
writing to it
each new fact rewrites the page: 7-15 s per fact on the hub, against about 1.2 s on small entity pages (server logs, during a 3,393-fact backfill)
page_versions
915 versions of that one page, 178 KB on average, 159 MB in all: a quarter of a 649 MB database. A person hub page: 653 versions, 58 MB
The rows already live in the facts table, and on a database-only source no Git file needs them in the body. Would you take an opt-in setting that leaves the body unchanged and keeps facts in the facts table only? Or one that renders a bounded fence (newest N rows, or a count with a pointer to recall)? That would solve the reads, the write time and the version growth together.
Related: #8192 (get_page omit_facts), #8274 (page_versions grows quadratically on hub pages), #8313 (curating an overgrown entity page), and PR #6877 (capture routing and facts.page_cap, closed 1 Nov as more surface than wanted).
reacted with thumbs up emoji reacted with thumbs down emoji reacted with laugh emoji reacted with hooray emoji reacted with confused emoji reacted with heart emoji reacted with rocket emoji reacted with eyes emoji
Uh oh!
There was an error while loading. Please reload this page.
On a managed Postgres brain (gbrain 0.60.565, database-only source, write-through off), every remembered fact is added as a row to its entity page's
## Factsfence. On a hub page, the fence ends up as most of the page and most of the write cost:recallalready serves from the facts tableThe rows already live in the facts table, and on a database-only source no Git file needs them in the body. Would you take an opt-in setting that leaves the body unchanged and keeps facts in the facts table only? Or one that renders a bounded fence (newest N rows, or a count with a pointer to
recall)? That would solve the reads, the write time and the version growth together.Related: #8192 (
get_page omit_facts), #8274 (page_versionsgrows quadratically on hub pages), #8313 (curating an overgrown entity page), and PR #6877 (capture routing andfacts.page_cap, closed 1 Nov as more surface than wanted).All reactions