Repository navigation
Replies: 1 comment
|
Two comments on the closed issue #5645 didn't move here with its text, and both bear on this idea:
We have run this change in production since early October. PR #5646 can be reopened if that helps. |
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Moved from #5645, opened by @howardpark on 2026-09-28. Feature requests now live in Discussions > Ideas.
Problem
An entity page's
## Factsfence grows with every remembered or extracted fact, and nothing bounds it: superseded and forgotten rows stay in the fence as history, and live facts accumulate as long as the entity matters. The rest of the page can be kept small by ordinary means (compiled truth is rewritten, the timeline can be rolled onto archive pages, a big topic can be split into sub-pages), but the fence cannot be split, because rows attach to the page they sit on.The documented way for an MCP client to change a page body is still
get_pagewithinclude_content: true, thenput_pagewith the wholecontentandexpected_revision. So every prose edit on an entity page carries its entire facts table both ways, and it cannot leave the table out: a remote put that omits a world-visible fence deletes those rows (the #4554 behaviour, which is right for a caller that saw them).Measured on a managed Postgres brain (0.59.x), one person page with ~190 fact rows:
get_pagewithinclude_content: true: ~128 KB response, about 50k tokens, twice Claude Code's default 25k-token MCP tool-output cap. Even acontent-only read (feat(get_page): content_only — the get→edit→put round trip without reading the page twice #5615) is ~63 KB, at the cap.get_chunks(the reader suggested on get_page: single-page body can exceed MCP response token cap with no way to paginate #3411): one unpaginated response of ~90 KB, 36 overlapping chunks that include the fence rows, so it does not fit either.So an MCP client cannot edit the prose of this page at all, and the page gets bigger with every fact. #3411 was closed asking for usage data before revisiting large-body reads; this is that data, for the specific case where the large part is the facts fence.
Precedent
Takes already work this way: a remote read omits the takes fence, and
preserveProtectedTakeskeeps the stored fence when a remote put arrives without it. Facts get only the row-level version (non-world rows restored), so a world-only fence always travels.Proposal
get_pagegainsomit_facts(boolean, default false). The facts fence incompiled_truth,timelineandcontentis replaced by a one-line placeholder,<!--- gbrain:facts:omitted -->.put_pagereplaces that placeholder with the stored fence, byte for byte, at the same position.expected_revisionalready guarantees the stored fence is the one the caller omitted.invalid_params.Facts keep changing through
remember/forget; only page edits stop carrying them. A PR in this shape follows.All reactions