Made what members say in the briefing reach chat (deal-breakers, goals, story)

Situation

After the onboarding briefing, chat told members it had nothing on file about their deal-breakers or what excites them, though they had just given both. The pitch used them, so members saw the product contradict itself. The cause was systemic: the memory extractor ran once per conversation, so in the briefing only the first interview turn ever became memory. Everything after it (career story, strengths, networking targets, next role, deal-breakers) never reached chat.

What I did

Investigated with the local prod copy: showed the agent reported its tools accurately and that the answers lived only in the briefing's progress tracker, which chat can't read. Quantified the gap across all briefings (below).
Chose to extract the briefing in windows (one per closed service area plus completion) over exposing the tracker's facts to chat, which would have created a stale second copy.
Built it with a cursor, a token-held lease, pay kept out of memories, and the chat-ingestion tool no longer silencing the extractor; routed the beacon/admin extraction through the same windows (it used to re-read whole briefings and could write pay as a memory).
Wrote a backfill for past briefings; ran three independent review passes; validated with evals, the ticket's exact chat question, and before/after on the QA accounts.
Added scripts/find-member-fact.sh to the workspace so "where did this fact go" takes one command.
PR: https://github.com/Qwestly/qwestly-agent/pull/481 · Ticket: https://app.asana.com/0/0/1219203102588930/f

Result

Pending: ships with qwestly-agent#481, then the production backfill. Locally, the QA account went from "I don't have explicit deal-breakers on file" to listing all four deal-breakers and what excites them.

Metrics

Metric	Baseline (2026-10-08, local prod copy)	Target
Completed briefings whose interview was extracted past its first turn	4 of 44 had more than one extraction batch (36 one batch, 4 none)	all new briefings, past ones via backfill
Members with a deal-breakers answer who have a memory resembling it	9 of 29	29 of 29 after backfill
Backfill scope (local copy)	39 completed briefings due, 587 member turns	0 due after apply

How to measure

Batches per briefing: for each onboarding_state with phase/step completed, count distinct created_at among user_memories with that source_conversation_id and source_type != "explicit".
Deal-breakers coverage: members with progress.service_areas.finds_opportunities.assessment.dealbreakers vs those with a user_memories row matching /deal.?breaker|won.?t|avoid|not interested|never/i.
Backfill scope: qwestly-agent/scripts/backfill_briefing_memories.py (dry run) under the shared env.
Run against a fresh prod copy (restore-prod-db skill) for production numbers.

Log

2026-10-08: baseline taken on the local prod copy; fix built and validated; PR #481 opened.