AIGuardian: A Parent-Regulated AI Chatbot Interface for Children

Abstract

Conversational tutors built on large language models (LLMs) personalize instruction effectively, but giving children unsupervised access to them raises three linked problems: exposure to confidently wrong or inappropriate content, the exclusion of parents from the learning loop, and the absence of durable memory across sessions. We present AIGuardian, a parentregulated tutoring system built on a real-time serverless backend (Convex) and a multi-lane agentic memory layer derived from the open-source Memobase engine. Memory is partitioned into Learner, Guardian, and System lanes held in isolated projects, so that a student's self-reports, a parent's directives, and productgenerated signals remain separate, attributable sources of truth. A guardian oversight loop distills lane-aware memory into short WhatsApp digests instead of exposing raw chat logs. We evaluate the deployed pipeline on a planted-fact benchmark of 15 synthetic learners (75 tutoring sessions, 1,800 turns). Buffered extraction reduces memory-processing tokens by 87.0% and cost by 88.1% relative to per-turn extraction, and flush cadence forms a measurable cost-quality frontier: an intermediate policy retains 66.4% question-answering accuracy from a ~1.7k-token memory context against 93.8% for a full-transcript oracle. Whenever both sides of an engineered learner-guardian conflict reach memory, lane attribution is preserved in 100% of cases. Audited claim by claim, 75.9% of digest claims are supported by the memory shown to the summarizer and only 3.5% contradict it, with no digest reproducing private conversational content.