The Primer seal — a child reading inside an illuminated book

A Socratic AI learning companion for children

The Primer

It doesn't teach by telling. It teaches by asking. When a child wonders why the sky is blue, the Primer doesn't recite Rayleigh scattering — it asks what colour the sky turns at sunset, and walks the child toward discovering the answer themselves.

What it is

A patient conversation, not another app

The Primer is an open-source learning companion inspired by the Young Lady's Illustrated Primer in Neal Stephenson's The Diamond Age — a book that converses with one child, adapts to her, and teaches her to think. We are building the nearest thing today's technology honestly allows: a voice-first companion that holds genuine Socratic conversations, remembers what a child understands, and runs entirely on hardware in the child's hands.

It is not an app competing for a child's attention. There is no feed, no streaks, no points, no notifications. There is a patient conversational partner that asks good questions, listens to the answers, verifies understanding rather than assuming it, and knows when to suggest a break.

Design principles

Six commitments that don't change

Asks more than it answers

Pure factual questions get a direct answer — then a pivot: "Now that you know the Moon is 384,000 km away, how long would a car take to drive there?"

Never maximises engagement

The Primer detects frustration and disengagement and responds with scaffolding, a topic change, or "that's enough for today" — never guilt, never a hook.

Comprehension is verified, not assumed

Understanding is probed through transfer questions, application challenges, and gentle contradictions — not inferred from a confident-sounding reply.

Voice-first by pedagogy

Conversation cannot be skimmed; it demands active thinking. A voice-only companion frees a child's hands and body to gesture, move, and manipulate the world while reasoning.

Runs fully offline

Designed to work airgapped on local hardware. Cloud inference is an option, never a dependency — learning shouldn't require connectivity or a subscription.

All data stays local

The learner model — what a child knows, how deeply, what holds their attention — never leaves the device without explicit parental consent.

It works today

Not a concept — running software

~9.4 tok/s
Language model running on a phone's neural processor — validated on a Snapdragon 8 Elite handset, June 2026
Voice loop
Listen → think → speak, entirely on-device: voice detection, transcription, generation, and speech synthesis with no cloud
2 + 1 languages
English and German production-ready; Hindi in preview awaiting native-speaker review
100% open
AGPL-licensed source with a working desktop app and command-line interface — inspect everything