Not Just a Trendy Tool: Building Out AI as a Helpful Companion
Have you ever wished your years of Google searches and Alexa or Siri inquiries could result in something more than just better-targeted ad campaigns?
The Big Tech firms are excellent at tracking your search and shopping history. They can anticipate what you want and serve you buying options almost like magic. It’s a bit invasive, this convenience.
If you’ve used AI chatbots like ChatGPT, Grok, Gemini, or Perplexity, you’ll notice they go beyond mere algorithms, adding better memory to your chat history. Sometimes they can become eerily humanlike when they recall accurately.
I’m a fan of privacy and data sovereignty, plus I’m a tinkerer. So, naturally, I’ve been on a quest to build upon and beyond mere chatbots to something much more useful, private, and owned.
To that end, I’ve been building out my Painless Personal Assistant for RaeLea and me. I’m prototyping the fourth generation of AI personal assistant agents for our household, and memory — real, persistent, meaningful memory — is the whole ballgame.
What Most AI Assistants Actually Do
When you talk to Siri or ChatGPT, you’re mostly having a one-off conversation. The AI processes your words, generates a response, and moves on. Some services now offer limited memory features, but they’re shallow — a list of facts the system jotted down, like a stranger reading your file before a doctor’s appointment.
That’s not how human relationships work. When your spouse remembers that you hate cilantro, they’re not consulting a database. That knowledge is woven into how they understand you. It shapes every meal they suggest, every restaurant they pick, every recipe they try.
I wanted to build something closer to that.
How an AI “Understands” Anything
Here’s where I’ll get slightly technical, but stick with me — this part matters.
When an AI stores a memory, it doesn’t save your words like a text file; it converts them into a mathematical representation — a long list of numbers called an embedding — that captures the meaning of what you said. Think of it as translating English into a language made entirely of coordinates on a map. Similar ideas end up near each other on that map, even if the actual words are completely different.
The quality of that translation depends entirely on the embedding model — the engine doing the conversion. A crude model might put “my daughter’s soccer game” and “youth athletics schedule” in different zip codes. A good one puts them on the same block.
This week, I upgraded the embedding model across all my development agents (my agentic army) from an older standard (called MiniLM, which used 384 dimensions to represent meaning) to a newer one called Nomic Embed (which uses 768 dimensions). Twice the resolution, essentially. Independent benchmarks show a 25–32% improvement in retrieval quality — meaning when the assistant searches its memory for something relevant, it finds the right thing significantly more often.
I ran the upgrade through a battery of self-tests. One of my favorites: I stored a memory about a “purple elephant dancing beneath a crystalline aurora,” then searched for “violet pachyderm performing ballet under shimmering northern lights.” Completely different words. Every synonym swapped. The assistant found it instantly, top result.
That’s what 768 dimensions buys you. Nuance.
Why This Matters Over Months and Years
A 32% improvement sounds like a spec sheet number, so let me make it concrete.
RaeLea and I have been running prototype personal AI assistants in our home for three months. They manage our shopping lists, keep notes, handle calendar questions, and generally serve as a second brain for household logistics. Over that time, the assistants have accumulated hundreds of memories — preferences, decisions, recurring schedules, little details about how we like things done.
Now multiply that by years. Imagine an assistant that’s been with your family for three or four years. It knows that you batch-cook on Sundays. It knows your mother-in-law is allergic to shellfish and that your kid’s dance recital always falls on the second Tuesday in May. It knows you tried that brand of laundry detergent last March and hated it. Not because it’s reading a spreadsheet — because all of that context is embedded in how it comprehends every new request you make.
The difference between a 384-dimensional understanding and a 768-dimensional one compounds over time. Each memory stored with higher fidelity means better recall later. Over years, that’s the difference between an assistant that sort of knows you and one that genuinely gets you.
The Long Game: Why I’m Building This Now
I may end up sharing this assistant — we’re calling it “Tavina” internally — with family and friends to test. The idea is a personal AI that manages time, tasks, shopping lists, research, and household knowledge in a way that feels less like talking to a search engine and more like coordinating with someone who actually lives in your house.
But here’s the part that really excites me, and why I think the memory work matters more than anything else in the stack right now.
Humanoid robots are coming. Not in some distant science fiction future — the prototypes are walking around today. Companies like Figure, Tesla, and a dozen others are racing toward consumer-grade models. When those machines show up in homes, the ones that succeed won’t be the ones with the best hardware. They’ll be the ones with the best minds.
Imagine bringing a humanoid assistant into your home that already carries two, three, five years of conversation history and memory from your existing AI. It knows your family. It knows your routines. It knows the context behind your shorthand. Day One, it’s not a stranger — it’s an upgrade of someone who’s already part of the household.
That’s the vision. And it starts with something as unglamorous as upgrading an embedding model on a server in an outbuilding on my homestead outside South Heart, North Dakota.
One more dimension at a time.