Preserving Childhood Beyond the Camera Roll
Why organising family photos is not enough: the product problem behind Childhood Archive, a private living archive for photos, voice and the stories a camera cannot know.
Follow on X ↗
· 8 min read
TL;DR: I started building Childhood Archive because my phone can contain thousands of photographs of my daughter and still do a poor job of preserving her childhood. The files are only part of the memory. The phrases she used, what made everyone laugh, who was there, what she was obsessed with that month and what her voice sounded like can disappear even when the JPEG survives.
There is a strange photo problem in modern families.
Previous generations could lose a childhood because too little was photographed. Today we can take tens of thousands of pictures and still struggle to find the moments that matter.
My own version of that contradiction is simple. I have a growing camera roll full of photos and videos of my daughter. When I ask my dad for photographs from my childhood, there is effectively nothing useful to pull out and revisit.
At first those seemed like opposite problems.
They are really the same preservation problem at two different scales.
A photo library is not a childhood archive
Photo libraries are very good at keeping media files. They can sort by date, search objects and faces, create albums and sync across devices.
But consider an ordinary photograph from a park.
The image might preserve the clothes, weather and people in the frame. Metadata might preserve a timestamp and GPS coordinate. A vision model might identify ducks, grass and a smiling child.
None of that tells you:
- the funny sentence she had just said;
- the name she used for ducks at that age;
- that it was the first afternoon after a difficult week;
- which grandparent took the picture;
- what song she kept asking for in the car;
- why an otherwise ordinary photograph mattered to you.
Those are exactly the details that become valuable after enough time passes.
This is the product gap I care about: how do you keep the media and the story without turning parenting into a documentation job?
The camera roll creates a second problem: fear of deletion
The more important a photo library becomes, the harder it is to clean up.
A phone can be almost full while a parent hesitates to remove anything because "it is in the cloud somewhere" does not feel like evidence that an irreplaceable original is safe.
That led to a non-negotiable rule in Childhood Archive:
upload success is not backup success.
The product should be able to distinguish at least these states:
- imported;
- exact original stored;
- original hash verified after storage;
- independent replica created;
- replica hash verified;
- eligible to remove the device copy.
That sixth state should not exist because a progress bar reached 100%. It should be derived from current evidence.
This sounds more technical than a family album should feel. That is intentional. The complexity belongs underneath the calm interface, not in reassuring copy that cannot be proved.
AI is useful here — but only with a boundary
AI can make a huge family archive useful.
It can cluster nearby photos into a likely moment, suggest that the same person appears across years, transcribe audio, identify duplicates, describe a scene and help a parent find "Christmas when she was four."
It can also make a permanent family record worse if its guesses are allowed to harden into history.
A model seeing a smile is not evidence that somebody "had the best day ever." A face match is not a relationship. The first photograph the system finds of a bicycle is not automatically a child's first bike ride.
So the archive uses a simple truth model:
- Observed — what media, metadata or a model suggests;
- Known — what a person explicitly supplies or confirms;
- Written — optional prose built from traceable evidence.
The principle is:
AI organises the memories. You tell the story.
Generated descriptions, embeddings and polished prose are disposable. The original media and human source context are not.
Voice deserves to survive too
One of the easiest things to underestimate is sound.
A five-year-old's voice changes. Mispronunciations disappear. The exact rhythm of a bedtime conversation becomes impossible to reconstruct from a transcript.
That is why I do not want voice notes to be a secondary attachment in this product. Raw audio should sit alongside photographs and writing as a first-class archival source. A transcript can make it searchable, but it cannot replace the recording.
The same applies to handwritten drawings, letters and little quotations. A childhood archive should not be constrained by whatever happened to fit into the camera roll.
Rediscovery matters as much as storage
A perfect archive nobody opens has solved only half of the problem.
The feature I am most interested in validating is called Tonight's Memories: a short set of a few photographs, recordings, quotes or prompts that a parent and child can look through together.
No infinite feed. No streak. No notification telling you that you have failed to document your child this week.
The point is to create a small ritual that makes the archive useful while childhood is still happening.
That also creates a natural moment to recover missing context. An old picture can prompt, "What do you remember about this?" A parent can answer in a sentence or record a short voice note. The archive becomes richer as a side effect of remembering together rather than a form that demands filling in.
Privacy changes the architecture
A family archive contains exactly the material that should not casually leak into logs, previews, search indexes or model-provider histories.
That means the public website and the private archive cannot be treated as the same SEO surface. The current fictional archive prototype is deliberately non-indexable. The public By JTT case study can explain the product; private family routes should not become search landing pages.
It also means ordinary analytics should not contain memory prose, transcripts, names, precise locations or filenames. Those fields would quietly create a second archive in a system built for operational telemetry.
And it means words such as "private," "encrypted," "backed up" or "safe to delete" have to be treated as engineering claims rather than marketing decoration.
What exists today, and what does not
The honest status matters.
The current project has a navigable fictional-data web experience and a substantial set of preservation, authorization, export, restore and deletion contracts. A current branch has deployed successfully on Vercel.
It is not yet a public family product. The dedicated backend still needs live isolation and recovery proof. Real authenticated import is not complete. The first-class Expo / React Native client for iOS and Android does not exist yet. Store-review builds do not exist yet.
I have now written the release destination down explicitly so the project cannot quietly confuse architectural progress with shipping progress:
- personal production;
- public web presence;
- native private beta;
- public App Store and Google Play release.
Each has evidence gates.
Why build it this way?
Because the archive is meant to matter more ten years from now than it does on launch day.
That changes the optimisation target.
I would rather have a boring export format than an impressive proprietary one. I would rather preserve exact original bytes than save a little storage by destructively transcoding them. I would rather say a feature is not enabled than display a fake privacy badge. I would rather keep AI output disposable than make a family dependent on today's model provider.
Most of all, I want to find out whether this solves the original problem in real life.
Can a large camera roll become less frightening to manage? Can the little stories stay attached to the pictures? Can looking back become something a parent and child actually choose to do together?
That is the test.
You can follow the build through the Childhood Archive project page as the private alpha moves from prototype to a product that has to prove what it claims.