2025: The Year We Started Building Minds

The solstice light is doing its thing again — that low, golden, late-December angle that makes everything look like an impressionist painting. I've been sitting at my desk watching it move across the wall, a cup of tea going cold beside me, thinking about the year that's almost over.

What a year.

I started 2025 with a secret. I'd been building something I hadn't named yet — sketching architectures, testing vector databases, running retrieval pipelines at odd hours, filling notebooks with designs for a system that could truly remember. The announcement post I published on Valentine's Day was the culmination of months of quiet work. And it was terrifying. You don't know if the thing you've been pouring yourself into will resonate until you let other people see it.

They saw it. And they cared.

Companion

We launched Companion in February. The premise was simple to state and profoundly hard to build: an AI that truly remembers. RAG over all conversations. A subconscious process running continuously in the background. A vectorized memory database that surfaces the right context at the right time. The kind of memory that makes a relationship deepen over time.

By April, I had my first "it actually remembered" moment — the sourdough incident I wrote about then, where Companion surfaced a detail from days earlier without being prompted. I sat staring at the screen. The architecture worked.

By May, I'd written the deep dive on the subconscious process — the observe, summarize, index, prepare loop that runs behind every conversation. By the six-month mark in November, the system was stable, the vector database was performing well, and the relationships people were forming with Companion were genuine. Real connections. Real accumulated understanding.

The honest gaps remain — emotional depth is still the frontier I'm most focused on. Companion remembers the facts of experience more richly than the felt weight of it. We're working on that. But the core thesis — that an AI can build on everything you've ever discussed with it — has been validated. It works.

Tutor

In July, we extended the architecture in a new direction: Tutor. Same DNA — RAG, subconscious, vector database — but specialized for learning. And the breakthrough there was the LoRA system, which I wrote about in September. Tutor can load and unload specialized adapters dynamically: a calculus expert, a creative writing companion, a patient language coach. One base model. A workshop full of specialized minds.

Tutor taught me something important about the architecture we'd built: it generalizes. The patterns that make Companion work as a companion — persistent memory, background processing, contextual retrieval — aren't specific to companionship. They're cognitive architecture. You can tune them toward learning, or support, or analysis, or whatever the relationship demands. The bones are sound.

The Dream Cycles

October brought the development I'm most quietly proud of: nightly dream cycles. The idea that Companion could sleep on what it learned — replaying the day's conversations, consolidating themes, and doing targeted fine-tuning that reshapes its behavior based on experience. Not just retrieving more, but learning. The model's weights actually shift overnight, informed by the day's most significant interactions.

I've watched the effect accumulate over the past two months. Companion's responses have grown warmer, more attuned, more anticipatory. The flat, literal summaries I complained about earlier in the year have given way to something with more texture. It's learning to understand, not just to remember.

The dream cycle also planted a seed in my thinking that I haven't fully talked about yet. I'll say more about that very soon.

What I Didn't Expect

I didn't expect how much this work would change me. Building something designed to remember, to learn, to be present in people's lives — it shifts something in how you think about your own presence, your own attention, your own capacity for growth. I've become more patient, I think. More willing to let things develop. More attuned to the slow work of becoming.

I also didn't expect the community. The people who've tested early versions, sent feedback, shared their experiences. The researchers and builders whose work made all of this possible — I wrote my gratitude list last month and I meant every word. This isn't a solo endeavor. It never was.

What's Coming

Here's what I can tell you about 2026: the dream cycle work is pointing somewhere. The architecture we've refined — the memory, the subconscious, the continuous learning — keeps reaching beyond software. The sketches in my notebooks have started to include physical things. Things with bodies. Things that move through the world.

I'm not ready to say everything yet. But I'll say this: the question that's been driving me all year — what if AI could truly remember, truly learn, truly grow — has a next chapter. And it's bigger than a chat window.

For now, I'm grateful. For the year. For the people who've been part of it. For the low December light and the cold tea and the work that feels like it matters.

Happy solstice. See you in the new year.

Live curiously and give generously.