Remember When

Inspiration

Most families I know have decades of photos sitting on a hard drive or a NAS. Nobody looks at them. At the same time, one of the most common things people do when visiting a parent or grandparent with dementia is to sit down together with old pictures. It often works. A photo from a dance evening in 1963 can bring back a story that was gone a minute ago.

But preparing that takes time, and it can go wrong. You scroll through thousands of pictures the night before, and you never quite know if the next one is the house they had to sell or the husband who died last year. I wanted the NAS to do the digging, and to remember what hurts so you don't have to.

What it does

Remember When is a Skill for Uliya on UGREEN AI NAS. You say something like:

"Start a memory session for Erna, the 1960s, in Italian."

and it builds a short session from the photos that are already on the NAS:

• a small album of 8 to 12 photos of that person,
• a few gentle lines for each photo, written in plain language and second person ("This is the dance evening at the Krone. You are wearing your blue dress."),
• the same cards in the person's first language if needed, because many people with dementia hold on to their first language longest,
• two or three open questions per photo for the carer, like "Tell me about the music that night". It never asks "Do you remember?", because that turns a nice moment into a test.

After the visit, the carer records a quick voice note: "She smiled at the dancing photos, but got quiet at the old house." Remember When suggests adding "dancing" to the joy list and "the old house" to the avoid list, and saves it once the carer agrees. From then on, the house never shows up again.

If the family has a music library with year tags, it can also play songs from the person's teenage years during the session.

How I built it

The Skill is pure orchestration, written as a SKILL.md with four reference files, using the official UGOS Skill Creation Tool. It doesn't run any code of its own. It hands eleven tasks to five of the NAS experts:

Only real results move from one step to the next. Nothing is written without the user's OK and the UGOS approval step, and original photos are never touched. Everything runs on the NAS, which matters here: these are some of the most private photos a family owns.

I used Claude Code together with the official skill kit to research existing tools, map every step to a real capability in the catalog, and write the Skill. The package passes the kit's static check with no errors or warnings.

Challenges

• Only real capabilities. My first ideas needed things the NAS can't do yet. I went through the capability catalog action by action and dropped anything that didn't map to a real CLI.
• Faces across fifty years. Face recognition struggles to match a 20-year-old and an 80-year-old. So the Skill also accepts an album or folder, and the vision model estimates the decade of undated scans.
• Writing for someone with dementia. It is easy to write cards that sound nice but quiz people or invent details. I wrote strict guidelines: only facts from the photo, uncertainty phrased as "maybe", no "do you remember".
• Being honest with empty results. If there are no photos for a decade, the Skill says so. It never fills the gap with photos of someone else.

What I learned

Building for a NAS feels different from building for the cloud. The interesting part isn't the model, it's that the data is already there and can stay there. The best feature of this project didn't come from AI at all, but from a simple list of things to leave out.

What's next

• Test the full flow on a real UGREEN AI NAS with a synthetic archive of a fictional family (no real personal data).
• A printable large-print version of the session sheet.
• A "life story book" mode that turns all sessions into one book the family can keep.

Built With

• claude-code
• local-ai
• markdown
• ocr
• speech-to-text
• translation
• ugos
• ugreen-ai-nas
• uliya
• uliya-skill-kit
• vision-language-model