Building MemoryLane: AI Memory Support for Families Facing Dementia

Some projects start with a market analysis. This one started with a song on the radio.

In March I was driving and Alan Jackson's "Remember When" came on. I had just spent weeks building a memory system for an enterprise AI platform — the kind that lets an assistant recall facts about you across sessions. And it hit me: I had been teaching software to remember people, while people in my own family were losing the ability to remember each other.

Alzheimer's has touched my family. I'll keep the details private, but anyone who has watched it happen knows the specific cruelty of it. The person is still there. The stories are still in there, somewhere. What goes first is the retrieval — the ability to reach back and pull a memory forward on demand.

That's a retrieval problem. I build retrieval systems for a living. So I started building MemoryLane.

## What MemoryLane Actually Does

MemoryLane is a memory-support app for families dealing with Alzheimer's and dementia. The core loop is simple.

A caregiver — usually an adult child or spouse — creates a profile for the patient. Then they start uploading memories: a title, a short story, an approximate date, the people involved, and a photo if one exists. "The summer we drove to the lake house, 1987. Dad insisted on towing the boat himself."

Then comes the part I think matters most: the caregiver doesn't do this alone. They invite siblings, cousins, old friends, and neighbors into what I call the Family Circle, using tokenized invite links. Each person contributes the memories only they hold. Your sister remembers the wedding differently than you do. An old coworker remembers a whole decade you never saw.

Finally, there's a Recall page. The patient — or a caregiver sitting beside them — can search those memories by typing or by voice. Ask about the lake house, get the lake house. Ask about a person, get every memory that person appears in.

That's the MVP, and it works end to end today: auth, patient setup, memory upload, contributor invites, and recall search. It's a demo, not a clinical product — I'll come back to that distinction.

## Why a Memory Graph, Not Just a Search Box

The naive version of this app is a photo album with a search bar. I wanted something closer to how memory actually works: associative, not just lexical.

Memories in MemoryLane live in Supabase — Postgres with the pgvector extension. Each memory gets embedded, so recall queries can match on meaning rather than exact words. Someone asking about "the boat trip" should surface the lake house story even if nobody typed the word "boat" into the title.

On top of that sits a memory graph: entities and relationships stored in their own Postgres tables. When a memory comes in, a serverless function (a Netlify function calling Claude Haiku) extracts the entities — people, places, events — and links them. Haiku is the right tool here: the extraction task is small and well-defined, it runs on every upload, and a heavyweight model would be wasted on it.

The graph is what makes the experience feel less like search and more like reminiscing. People and places become first-class nodes connected across memories. Pull one thread and related ones come with it. That mirrors something real about dementia care: clinicians have long used reminiscence work, where one recalled detail cues another. The graph structure is built to support exactly that kind of associative walk.

The other piece in the vision docs is AI-generated illustrative images for memories that have no photo. A lot of the most precious memories — especially older ones — were never photographed. A gentle illustration of "the kitchen on Sunday mornings" gives the recall experience something to anchor to visually. That's on the roadmap rather than in the MVP, but the schema is designed for it.

## The Family Circle Is the Product

Here's the thing I got wrong in my first sketch: I designed for one caregiver entering data. That model fails in practice, because no single person holds a life's worth of memories — and the primary caregiver is usually the most exhausted person in the family.

The Family Circle flips it. Memory collection becomes distributed. The daughter in another state, the brother who only saw Dad at holidays, the friend from the old job — each contributes from where they are. The burden drops and the corpus gets richer at the same time.

It also does something quieter. Contributing a memory is a way for far-away family to participate in care when they can't be physically present. People want to help and don't know how. "Write down three stories about your grandfather" is something everyone can do.

Technically, the invite flow took real care. Invites are tokenized links, and the token redemption runs through security-definer RPCs in Postgres — `get_invite_by_token` and `accept_invite_by_token` — so that row-level security stays strict on the underlying tables while still letting an invited person who doesn't have an account yet redeem their link. RLS policies for a multi-party app like this are genuinely fiddly; my last commit message on the repo is literally "fix: rewrite RLS policies." Health-adjacent data and family permission boundaries are exactly where you want the database, not the application code, enforcing who sees what.

## Voice-First, Because the User Is 80

Most software is designed by 30-year-olds for 30-year-olds, and it shows the moment you watch an 85-year-old try to use a hamburger menu.

The patient-facing side of MemoryLane is designed voice-first. Typing is a barrier — arthritis, tremor, unfamiliarity, or the disease itself. Speaking is not. "Tell me about the lake house" is a sentence a person with moderate dementia can still produce long after they've stopped being able to navigate an app.

Where there is a visual UI, the spec targets WCAG AAA, not the AA most products settle for, with 64px touch targets — roughly double the common minimum. High contrast, big type, no nested navigation, no clever gestures. Every screen should answer one question: what is this, and what's the one thing I do here?

This is also why I think modern AI changes what's possible in this category. The hard part of elderly-accessible software was never the backend. It was the interface. Speech recognition and language models good enough to hold a natural conversation remove the keyboard and the menu tree entirely. The interface becomes talking — the one interface this generation has used their whole lives.

## Why This Market Is Underserved

Dementia care tech is a strange market. The need is enormous and growing with the demographics, the emotional stakes couldn't be higher — and yet the consumer software shelf is nearly bare. Most of what exists targets institutions: facility management, staff scheduling, clinical documentation. Families at home get pamphlets and a support-group URL.

I have a theory about why. The buyers are stretched-thin caregivers, the end users have declining abilities, and the outcomes are hard to quantify — so it doesn't pattern-match to what venture-scale software looks for. The startups that try usually drift toward the institutional buyer, because that's where the predictable contracts are.

But the funding picture for the science is real. While researching this space I found that NIH SBIR grants for Alzheimer's and related dementias run up to $500K — non-dilutive money specifically aimed at this problem. That tells you the public-health side considers it urgent even where the consumer market hasn't caught up. The gap between "validated clinical need" and "anything a family can actually download" is the opportunity.

And to be clear about where MemoryLane sits: it's a working MVP, not a medical device. Invite links get copied by hand, the storage buckets aren't locked down for production, and it runs on sample data. I'm explicit about that in the repo. Anything that touches a clinical claim needs evidence, review, and a much higher security bar. The current build proves the experience; the research folder next to it holds the PRD and the clinical literature for what comes after.

## The Stack, Briefly

For builders who want specifics: Next.js 14 with the App Router and TypeScript, React 18, Tailwind, framer-motion for the gentler transitions. Supabase carries the heavy load — Postgres with pgvector for embeddings, Auth, Storage, and RLS for the permission model. Netlify handles deploys and the serverless entity-extraction function. Nothing exotic. The novelty is in the data model and the audience, not the infrastructure.

## A Practical Takeaway

If you're building AI products for elderly users — or for any population the industry usually ignores — here's what MemoryLane taught me:

1. **Design the data model around contribution, not entry.** One overloaded admin typing everything in is how these tools die. Make it trivially easy for many people to add a little.
2. **Treat voice as the primary interface, not a feature.** For users who can't or won't type, voice-first isn't a nice-to-have. It's the difference between a product and a demo.
3. **Use a graph plus embeddings, not embeddings alone.** Vector search finds similar text. The entity graph is what lets one memory cue another the way human recall actually works.
4. **Put permissions in the database.** Multi-party family data with RLS and security-definer RPCs is harder up front and much safer forever after.
5. **Exceed the accessibility spec.** AA is a floor written for the general population. If your users are 80, build to AAA and oversize everything.

The deeper lesson is simpler, though. The most meaningful application of AI memory I've found so far isn't making an assistant smarter. It's helping a family keep hold of the stories that made them a family — and making sure that when someone asks "remember when?", there's still an answer.