How the agents learn
Six pipelines exist to make replies better over time. This page is the honest scoreboard: what each one does, whether it genuinely works today, and what stands between the broken ones and life. Audited against live data — not the design documents.
flowchart LR
EP["📼 Episodes
every bot turn recorded"] ==>|"✅"| SC["Nightly scorer
did the turn work?"]
SC ==>|"✅ fixed Aug 2026"| UI["Per-customer
behavior notes"]
UI ==>|"✅ refilling"| P["🧠 the agent's prompt"]
KB["📚 Company knowledge
authored policies"] ==>|"✅"| P
AM["🧠 Agent memory
one note per customer"] ==>|"✅ feeds BOTH agents"| P
QB["❓ Question Bank
29 questions"] ==>|"✅ serving ON — grows with approvals"| P
CORR["📝 Staff corrections"] -.->|"⚠️ ready, never used"| P
AN["📊 Analyst signals"] -.->|"❌ frozen since Feb"| P
P ==> REPLY["💬 better replies"]
classDef ok fill:#e7f2ec,stroke:#166b4e,color:#182420
classDef idle fill:#f7edd8,stroke:#b4711a,color:#182420
classDef dead fill:#f6e3de,stroke:#a63a2b,color:#182420
class EP,SC,UI,KB,AM,QB ok
class CORR idle
class AN dead
Solid arrows flow today. Dashed ones are built but not yet feeding anything.
The scoreboard
| Pipeline | What it does | Status |
|---|---|---|
| Company knowledge | Hand-written policy pages (prices, rules, services) retrieved into the agent's prompt as the authoritative source it must never contradict. | WORKING |
| Agent memory | One short note per customer, refreshed after a turn only when something worth remembering was said — facts no other table already owns. Read by the conversation agent, by the outreach planner when it decides whether and what to send, and shown to staff in the Dashboard. It is the one thing written by one agent and read by the other — the real bridge between them. (A second, unrelated read of it inside the outreach signal job was deleted on 7 Aug; the planner's read is untouched.) | WORKING rebuilt 7 Aug — currently empty, refills on the next substantive turn |
| Episodes → outcomes | Every turn is scored at night: quick reply? repeated question? went silent? completed a profile after? Scores roll up into per-customer behavior notes injected into future conversations. | WORKING repaired Aug 2026 — was silently dead for months |
| Question Bank | Deduplicated customer questions with staff-approved answers, served on meaning-match. Full page → | SERVING ON |
| Staff corrections | When staff correct a bad bot reply, every future similar situation shows the agent "don't repeat this mistake — prefer this wording." Fully built and waiting; no correction has ever been submitted. | READY, UNUSED |
| Analyst signals | Conversation-quality metrics from the analyst service. Its writes have been silently failing since February — known root cause, queued for repair. | FROZEN |
Why the episode loop matters most
It's the only pipeline that learns from outcomes rather than from what someone wrote down. A repeated question is a reply that failed. A customer who goes silent after a message is a signal. A profile completed the day after a conversation is a win worth attributing. For months these signals were computed by a job that crashed nightly while reporting success — the fix (August 2026) drained the entire backlog in one run, and two outcome signals fired for the first time ever the same night.