How Minie mapped 103 memories about a 4-year-old in 3 months

The other AIs remember — just not for your kid

ChatGPT remembers. Gemini remembers. Claude is starting to. The problem isn't that they've forgotten how to remember — it's that their memory is built for a single adult user's life. When a kid uses them, that memory works against you.

• Shared adult account. Your 4-year-old asks a question on the family iPad; it's logged in as you. Favourite bedtime stories and your quarterly deck land in the same memory bucket. Context bleeds both ways.
• Built for adult facts. General AI memory optimises for what adults reference explicitly — preferences, project history, writing style. Kids need the opposite: how they refer to a story, who they mean by "Dada," the made-up word for penguin.
• No safety filter. Adult AIs cheerfully remember anything. If a 4-year-old offhandedly discloses something sensitive, it's stored, indexed, and can be surfaced back weeks later. Kid-first memory has to refuse to remember some things.

From hundreds of conversations to 103 structured facts

Every conversation passes through an extraction pipeline. Messy turns in — interruptions, language switches, half-finished thoughts. Structured facts out:

• "Aadya's favourite story is Under My Bed."
• "Aadya can count to 25 and backwards."
• "Aadya speaks Hindi, English, Tamil."
• "Aadya's dad is called Dada."

Each fact is:

• Entity-tagged — Story, Person, Place, Skill, Language.
• Cluster-linked — a princess-dragon story belongs to both Stories and Magic.
• Deduplicated — repeats strengthen the existing fact, not a new one.
• Contradiction-aware — "favourite colour is blue" → later "red" updates, not duplicates.

Why traversal beats search for a 4-year-old

The graph isn't cosmetic. It's why Minie gets smarter. Without it, every phrase is a cold keyword search. With it, Minie has context — the way a human remembers you.

"Play the bed one."

Without a graph: the AI hears "bed" and "one" and searches dumbly. With a graph: Minie finds the Stories cluster, sees Under My Bedhas been played six times, knows "bed one" is how Aadya refers to it, and plays it.

"Dada said no."

Without a graph: "Who is Dada?" — jarring for a 4-year-old. With a graph: Minie already knows Dada is her father and responds naturally.

"The space one from yesterday."

That session's story is linked to the Space cluster. Found in milliseconds. No five-round clarification.

30 → 60 → 103: how the graph compounds

Aadya's graph started with a handful of memories after her first sessions. Three months later: 103 memories across ten clusters. Every conversation adds a few more — another story titled, another pet named, another favourite planet.

Because Minie reads the graph at the start of every session, the conversations get more personal as the graph grows. Month one: Minie asks what kind of stories your kid likes. Month three: Minie already knows she wants the rabbit story with the sad ending, not the happy one.

That compounding is the point. The longer your child uses Minie, the more the graph looks like them — and the more the responses feel like they were written for your kid specifically. Because they were.

Privacy — three pillars, no handwaving

1. Sensitive data never enters our system

Before a single fact is stored, the transcript and extracted memories pass through a safety filter. These categories are blocked at extraction time and never persisted:

• PII — phone, email, address, full legal names
• Trauma disclosures — abuse, violence, self-harm
• Negative self-labels — "I'm stupid," "nobody likes me"
• Family conflict — "parents fight," "mom cries"

These are handled warmly during the conversation — a gentle redirect, a parent-notification if severe — but they're deliberately not stored. Re-surfacing them would risk compounding distress. Forgetting is a safety feature.

2. Access-controlled storage, with honest review while we're small

Every memory is keyed to an opaque child UUID — the name "Aadya" appears nowhere in the index. Memory is stored in an access-controlled MongoDB Atlas database with AES-256 encryption at rest. Access is bound to a worker service account; database credentials live in a separate restricted store.

At our current scale — two full-time founders, ~100 users — our team reviews conversations and extractions manually to catch bugs, tune the safety filter, and improve how Minie understands kids. This is how every early-stage AI product works, and it's how we make the thing good. Reviews are logged and restricted to the founding team. Per-child application-layer encryption is on the roadmap as automation replaces manual review.

3. You own the graph

Today, you can already delete all your data or any individual conversation from inside the Hey Minie app — associated memories clear from the graph within 24 hours. The graph viewer, single-memory delete, extraction pause, and JSON export are launching in the next Hey Minie major release, bundled into the existing Hey Minie premium subscription.

"We built Minie's memory the way we wanted it for our own kids. What the graph refuses to remember matters more than what it does."
— Manjush, co-founder

Your child's graph will look nothing like Aadya's

That's the point. Every graph is a fingerprint of a specific kid's mind — the stories they pick, the questions they ask, the people they love. Aadya is a founding-team child, shown with her family's written consent. No real user's data is ever made public.