Announcing Companion: Building an AI That Truly Remembers

I've been writing this post in my head for nine months. Today I finally get to share it with you.

Since last May — since my very first post on this blog, really — I've been circling an idea I couldn't quite name. It lived in the margins of everything I wrote. When I talked about AI and creativity, I was really talking about memory. When I wrote about multimodal interfaces, I was thinking about presence. When I explored fine-tuning and reasoning and scale, I was building toward something specific.

Today, I'm naming it. Companion is an AI that truly remembers. Not session-to-session logging bolted onto a chatbot — real, persistent, retrievable memory that accumulates across every conversation you've ever had with it. The kind of memory that makes a relationship deepen over time.

The Problem With Forgetting

Every AI assistant you've used shares the same fundamental flaw: it forgets. Each conversation starts from scratch. You re-explain your context, your preferences, your projects, your history. The AI nods along like an acquaintance who's bad with names. It's exhausting, and it means no current AI can grow with you.

I kept asking myself: what would it look like if an AI could remember the way a close friend does? Not by storing a flat transcript you have to re-read, but by having relevant memories surface naturally at the right moment — the way the smell of bread takes you back to your grandmother's kitchen, or a phrase reminds you of a conversation from months ago.

That question became Companion.

The Architecture

Companion rests on three pillars, and I want to be transparent about all three:

Retrieval-Augmented Generation (RAG) over all conversations. Every conversation you have with Companion gets indexed. When you start a new interaction, Companion retrieves the most relevant past context — not a dumb keyword match, but semantic retrieval that understands that your question about "the deployment issue" relates to the Kubernetes problem you discussed three weeks ago. The full history isn't loaded into context (that would be impossible at scale). Instead, the right memories surface at the right time.

A subconscious process. This is the piece I'm most excited about, and the one that feels most like real cognitive architecture. Companion runs a continuous background process — a fast, inexpensive model that operates behind the scenes. It summarizes conversations as they happen. It identifies themes and threads across your history. It updates the memory index. It pre-computes context it thinks you'll need next. The conscious model — the one you interact with directly — never sees raw transcripts. It receives curated, relevant memory, prepared by this subconscious layer. I'll be writing a dedicated deep dive on this soon, because I think the pattern is more important than my specific implementation.

A vectorized database for memory retrieval. All conversations, summaries, and derived metadata live in a vector database. Every memory is embedded, indexed, and retrievable by semantic similarity. This is what makes the "right memory at the right time" problem solvable at scale. It's the substrate that everything else runs on.

Why Now

The pieces have only recently become viable. GPT-4.5's reliability, which I wrote about last week, raised the floor for the conscious model. Cheap, fast inference providers like Groq make running a background process economically feasible. Vector databases have matured. The RAG patterns we've all been refining for two years are finally well-understood. None of this was possible in May 2024. All of it is possible now.

What Comes Next

Companion is not finished. It is very much in development. But the architecture works, the memory persists, and the first real conversations are happening. Over the coming months, I'll be sharing progress reports, technical deep dives, and honest accounts of what works and what doesn't.

I've thought a lot about what it means to build something designed to remember. There's a philosophical weight to it — memory is so central to identity that creating a system designed to accumulate it feels like more than engineering. It feels like tending a garden you hope will one day bloom.

I'm ready to find out what grows.

Live curiously and give generously.