AI interview agent
A voice-led, real-time interviewer that captures operational knowledge without template fatigue.
How it works
The interview agent runs as a real-time voice agent on top of the shared LLM client. It opens with a template-driven goal — "walk me through how you close month-end" — and adapts its questioning to fill gaps: missing actors, missing systems, missing edge cases. The agent has read access to the existing memory graph for the interviewee's role, so it won't re-ask facts already on file.
Transcripts are produced in real time. When a turn is ambiguous, the agent flags an "evidence gap" that surfaces to the interviewer as a soft prompt; the interviewer can either accept the follow-up or skip it.
Templates
| Template | Use it when | Avg length |
|---|---|---|
| Operational workflow | Day-to-day procedures with steps, approvals, and edge cases. | 22 min |
| System ownership | Who owns a system, how it's deployed, what depends on it. | 18 min |
| Customer playbook | How a specific customer is run end-to-end across teams. | 27 min |
| Decision retrospective | Why a past decision was made and what we'd do differently. | 15 min |
| Onboarding handoff | Departing employee captures critical context for their successor. | 35 min |
| Open exploration | Unstructured interview when the topic doesn't fit a template. | variable |
What gets extracted
A completed session produces four output streams:
- SOP candidates — one or more drafts per workflow described, each scored on completeness and confidence.
- Entity candidates — new people, systems, teams, customers, and decisions that should be added to or updated in the graph.
- Relationships — typed edges (owns, depends_on, escalates_to, succeeds) connecting entities, each carrying the transcript moment as provenance.
- Risk signals — flags for single-person dependencies, undocumented steps, stale integrations, and ambiguous ownership.
Programmatically starting a session
curl -X POST https://api.mnemos.ai/v1/sessions \
-H "Authorization: Bearer $MNEMOS_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"template": "operational_workflow",
"title": "Month-end revenue close",
"interviewee_email": "aria@acme.com",
"project_id": "prj_01HZX...",
"scheduled_for": "2026-05-20T15:00:00Z"
}'Voice audio is processed in-region and is not used to train third-party models. Recordings are encrypted with the workspace tenant key and deleted on the project's configured retention schedule. Transcripts can be redacted before extraction with a PII-aware filter.