Abstract
This defensive publication discloses a closed-loop memory architecture for real-time voice AI agents in which the machine-readable output of a post-call extraction step is fed back to reconfigure the next generative session. On call teardown, a single generative pass jointly extracts the spoken caller identity and composes a personalized follow-up email, halving per-call inference cost relative to running extraction and composition as two passes. A per-persona configuration row gates a five-way action fan-out — follow-up email, admin briefing, CRM contact, CRM lead/ticket, and caller memory — over that single result. CRM lead creation exploits a unique-constraint violation as a control signal: on collision the mechanism finds the existing lead by the extracted email and appends a timestamped transcript note under a rolling character cap, producing an accreting dossier rather than duplicate rows. A MERGE upsert keyed on normalized phone accumulates call count, topic strings, rolling summaries, and an IVR-captured preferred language. The next inbound call's greeting injector primes the live session with the caller's name, prior call count, prior topics, and a resolved language lock — injected during the greeting delay window. This document establishes dated, enabling public prior art over the mechanism and its constituent techniques.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Assuncao, gustavo matthew, "Closed-Loop Cross-Call Caller Memory for Voice AI Agents", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/10928