Person Briefs, Long-term Memory, and Corroborated Linking

Shipped since June 2.

Person briefs

A dossier for any person, by name (prospects included):

• Role: teammate or external prospect.
• Activity counts from git and meetings.
• Their decisions, concerns, priorities, and requests, with confirmed and pending-review listed separately.
• Each request linked to its initiative and current status.

Long-term Memory panel

• A panel listing the org’s interpretive claims, attributed to the person who said it.
• Confirm or dismiss each claim from a drawer.

Model selection

• The model you pick is used on every ask.
• On a mid-stream model overload, the request retries or falls back to another model.

Link corroboration

• Link proposals are checked against deterministic signals: explicit (INI-57) tags, pull request text, branch names, exact initiative names, and the files a commit touches.
• An explicit tag overrides inferred ownership and the date-window eligibility check.
• Merge commits are excluded from attribution.
• Ownership patterns are ranked by specificity.
• A daily audit reports the verified-tier counts, and reports that precision is unmeasured until wrong suggestions are dismissed.

New search tools

• Meeting search.
• Conversation thread search.
• Initiative status history.
• Per-team activity and merge-rate.
• Roadmap search.
• Fetch a specific URL or PDF; flagged when it returns no sources.

Meeting-transcript memory

• Extraction reads full transcripts, including plain-text meetings with no per-line rows.
• Long transcripts are chunked instead of truncated at 12k characters.
• A materiality filter drops to-dos, logistics, and chatter.
• Personal-scoped meetings are excluded from org memory.

Release

• Backend, frontend, and the model layer. Deployed June 5.
• 30+ merged pull requests.
• Running in production. Database changes are additive.