Solving the Problem of Remembering. Again and Again and Again. Until It Finally Worked.

Imagine you come to see me in the clinic but instead of normal conversation with you, I'm acting weird.

Before I answer or respond to a single thing you say, I reread your entire file - all our previous communications etc before I give you a response.

Not reading your notes before our appointment together. Reading them before being able to respond to every. single. sentence.

"How are you?" - furiously flicks through 36 pages of notes.

"I'm good thanks, how are you?"

•Gets ready to furiously flick through notes again•

That's how my chatbot Biggie remembered you at first.

It worked for a few responses. Until it didn't work anymore and Biggie forgot the topic of conversation, answered in a way that made little sense and needed to be prompted more than a goldfish would.

So then I figured it would be wiser to inject a summary of your onboarding and chat history to date into the pipeline so that Biggie was always aware of what we had talked about before.

This was much better, he wasn't forgetting nearly as much and could stay on topic and responses were contextual and specific to me for the most part.

What i hadn't realised is that I had jumped the lobster out of the boiling pot of water into the pot beside it that was heating up just a little bit more slowly.

The problem was that as I chatted with Biggie, the summary that Biggie was given access to had to change.

The file kept growing. So every reply had more information to process and the replies got slower. And more expensive.

No bueno David.

And the stuff that actually mattered got buried under the stuff that didn't. There was no way for my chat system to understand that the fall you had and recovered from 18 months ago wasn't relevant now.

Or the poor nights sleep you'd had last week actually happened last week and that you had shaken the effects of that off entirely.

I'd built a thing to cut through noise... that literally felt like it was standing in front of the speakers at Lollapalooza.

Mucho no bueno David.

---

So I'm chatting to Claude about it (spotting a theme yet?).

And after describing the problem and using my logic donkey skills (I love that phrase, someone described me as so strong with logic that I was a logic donkey & it's always stuck with me!). So anyway, out of our conversation falls something called Memory Augmented Retrieval.

Sounds fancy...

The basic premise is this:

Don't reread the whole set of notes every time. Just pull the right piece from the right page at the right moment.

It remembers everything and then surfaces only what's relevant to what you just said, what's recent and what actually counts.

Exactly what a good clinician does. Except that it doesn't forget like a human can do at times. And it can even flag contradictions and reconcile them across time.

Pretty useful for a support system for something as complex as chronic pain.

Goldfish memory be gone!

After that my chat with Claude gets a little more confusing because apparently this MAR thing leads to something called a user context graph.

I couldn't draw in primary school. Never mind building a graph for every single user in the back of my app.

What the hell is that about?

Turns out it's one of the next rabbit holes I will probably have to wander down.

So I wrote it on the list... and closed the laptop.

Some doors you walk through later I guess.