Repository navigation
Replies: 2 comments
|
+1. We're running into the same thing on a different setup, a shared brain rather than a personal one, so the proposed repair would help here as well. Setup: gbrain 0.60.728.0 (stable), Postgres, managed persistence. Observed in that shared source:
Why it matters on a shared brain: for us the source grant decides who can read what, and Things we checked:
What we'd need: the |
|
+1, from a personal managed brain on v0.60.613.0 (Postgres). About 14,000 active
A shape that would cover it: an owner-only
We run a local version of 1 to 3 (about 30 source lines plus a five-case PGLite test: opens the named rows and keeps every other column; leaves fenced, expired, already-world and private-page rows alone; a raw |
Uh oh!
There was an error while loading. Please reload this page.
Moved from #5857, opened by @andreineacsu on 2026-10-02. Feature requests now live in Discussions > Ideas.
Summary
#5671 fixed the default for new facts on HTTP-only installs: with
facts.default_visibility world, new facts are readable over MCP. Facts that were alreadyprivatestay unreadable to every MCP caller, and I could not find a documented way for the operator to widen them. On a personal brain that leaves most of the agent's facts reachable only through the trusted local CLI.They also do not age out. When a page is rewritten over MCP, its existing private
## Factsrows come back as new rows that are stillprivate.Environment
gbrain serve --http: onegbrain bootstrap harnesstoken (headlessclaude -pturns and scheduled jobs on the brain host) and one OAuth client (Claude Code on a second machine of the same operator)brain.audience personal;search.remote_private_pages visiblesince 2026-09-25;facts.default_visibility worldsince 2026-09-29, following step 3 ofskills/migrations/v0.60.20.0.mdObserved
Active facts by visibility:
Of the 10,260 private facts, 6,438 have no
source_markdown_slug(database-only) and 3,822 sit in## Factsfences on 282 pages. Over MCP,recall,entityandget_pageleave all of them out. The page opt-out does not reach them either:docs/operations/mcp-surface-runbook.mdsays remote page bodies keep only world facts.New private facts keep appearing after the default changed:
All 450 private rows written by
mcp:put_pagesince the change describe events from before it, while almost every world row is new. That is consistent with the reconcile re-inserting a page's existing private rows (which a remote put preserves) as new rows. As far as I can tell, the private set therefore never shrinks without an operator action. This may be related to the ID churn described in #5430.The trusted local CLI does return these facts (
gbrain recall <slug>). In a Claude Code agent its output lands in the same context window as the MCP results, so on this topology the private tier decides which transport the agent must use, not who can read the facts.What I looked for
docs/protocol/MEMORY_VERBS_v1.md:privatemeans local CLI reads onlydocs/operations/mcp-surface-runbook.md:search.remote_private_pagescovers pages, not fact rowsskills/migrations/v0.60.20.0.mdstep 3: sets the default for new facts onlygbrain repair --help: thevisibilitykind stamps atoms and synthesized concepts, tighten-only, and touches no factsI found no operator path to widen existing private facts.
Expected
One of these, for a brain whose operator declares it personal:
gbrain repairkind would fit: dry run by default,--sourceand--entityfilters, written through the coordinator so fences and the facts table stay in agreement, resumable.search.remote_private_pages, so remote callers on a single-user brain can read private facts when the operator opts in.If neither fits gbrain's model, a line in the docs saying that legacy private facts stay local-CLI-only by design would also settle it.
Related
#5671, #5605, #5430, #4546
All reactions