Adherence Memory Callback

Adherence Memory Callback

Use this skill when a pharmacy or clinic (with recipient consent) wants a brief outbound phone check-in on how a patient is getting on with a prescribed medicine — and wants the agent to get smarter with every call instead of starting from zero each time.

It packages a two-tier memory on top of a CALL-E outbound call:

• a sub-brain per caller (private running summary, open items, and any "call me back" context), and
• a shared master brain of general facts and anonymized signals learned across all callers, guarded so no single caller can poison it.

This skill only listens, acknowledges, and notes answers. It is not medical advice: it never diagnoses, never recommends a medicine or dose, and escalates anything serious to a human pharmacist.

When to use

• One outbound medication-adherence check-in to a consented patient number.
• You want per-caller continuity ("last time you mentioned…") and cross-caller learning (patterns several patients report).
• You want a human-in-the-loop gate before the agent starts proactively asking about a newly learned side effect.

When not to use

• Diagnosis, triage, dosing, emergency response, or any medical advice.
• Unsolicited outreach, marketing, or lead generation.
• Recurring schedules without a separate scheduler wrapper and explicit consent.

Workflow

1. Read references/safety.md and confirm recipient consent and that it is not quiet hours for the caller's region.
2. Build the call goal from memory: the caller's sub-brain (continuity + any callback context) + the master brain's canonical facts (background) + any admin-approved proactive questions + the safety rails.
3. Preview first (no call): run the reference app in --dry-run mode to see the exact goal.
4. Place the call through CALL-E only after consent and guard checks pass.
5. After the call, extract structured fields from the transcript and update memory: the sub-brain summary/open items, candidate facts (through the corroboration gate), and anonymized signals.
6. If the caller asked to be called back, store the short reason so the next call opens with it ("last time you were at a wedding — how did it go?").

The corroboration gate (why this is safe)

A learned fact stays a candidate until at least two distinct callers independently report it — the same caller repeating themselves never counts. Only then does it become canonical and eligible to influence future calls. See references/safety.md for the full curation and privacy rules, and references/examples.md for worked call-to-memory examples.

Human-in-the-loop prompt approval

A pattern reported by enough distinct callers is surfaced to an admin, who confirms or dismisses it before the agent starts proactively asking about it. A strong signal (many distinct callers) can auto-apply; the admin can also require manual approval for every change. The agent asks a proactive question only about approved patterns, and if the caller says they have not had the issue, it reassures them and tells them to contact the pharmacy if they ever do.

Runnable app

The reference runner lives at apps/python/cortex-call-brain/ (relative to this submission repository root). It builds the CALL-E goal from memory, applies consent / quiet-hours / idempotency / budget guards, places the call through the CALL-E CLI when run live, and folds the transcript back into the brain. Use its --dry-run mode to preview a goal without placing a call.

Output

After a completed call, expect structured fields such as:

• outcome — one of adherent, side_effect, needs_refill, missed_doses, no_answer
• sub_brain_summary — a short private summary of this caller for next time
• open_items — follow-ups for the next call
• candidate_facts / signals — general, anonymized knowledge for the master brain

Mask phone numbers in any user-facing summary.