The story

The version I wanted didn't exist.

Personal AI is being decided right now, and mostly by companies whose pitch is the same shape — hand us your conversations, your memories, your work, and you'll get an assistant back. Your stuff lives in their cloud. The convenience is real, and so is the price: you become legible to a system whose interests aren't yours.

I wanted the other shape. Local-first, on hardware I own, with memory that doesn't leak. The same things you'd expect from a paid product — chat, narrative, research, agents, coding — but mine. Not rented.

So I started building it in March 2026. By August 2026, this is the shape it has grown into.

What I was trying to escape

The cloud assistants are very good. That part isn't in dispute. What I couldn't shake was the slow accumulation: every conversation, every memory, every draft, every reflection — landing in a vendor's warehouse, indexed against me, used to train the next version, subject to a privacy policy that could change next quarter. The product I was getting was real. The shape of the relationship was wrong.

It's not just one provider. The whole posture of consumer AI right now is "trust us, we'll be careful." Maybe they will. They've also shipped features, deprecated them, raised prices, sold to investors whose interests aren't the user's, and rotated through governance structures faster than any of us can track. Trust at that scale and velocity is a bad bet.

The other thing I noticed: I couldn't actually own any of it. Not my conversations. Not the memories the model had built about me. Not the workflows I'd spent months tuning. When a model deprecates, the relationship I had with it goes with it. Years of context, evaporated in a release note.

The version I wanted

I wanted something that lived on my hardware, that remembered me across sessions because I kept the memory, that I could rebuild from scratch if something broke, that didn't depend on any company's continued goodwill to keep working five years from now.

I wanted it to handle the things I actually use AI for — quick questions, long-running stories, research with citations, agents that ship deliverables, coding help with a workspace it actually understands. Not five tabs in five different products with five different bills.

I wanted privacy as the default, not a paid tier. I wanted cloud providers to be opt-in, not opt-out. I wanted the off button to actually mean off.

And — this is the harder part — I wanted it to be good enough that I'd actually use it. Local-first doesn't help if the product is worse. A privacy-respecting tool nobody wants is just a tool nobody uses. So the engineering had to be real.

The first stretch

The early burst was real, and the scope reads strangely against that timeline. Some of it is genuine engineering, some of it is using Augmentum to build Augmentum, and a lot of it was focused launch-window intensity. The pace will slow; the product needs polish more than spectacle now.

Augmentum started as a proxy between LLM frontends and backends — a place to add the smart routing the existing tools didn't have. Then the proxy needed memory, and the memory needed a substrate, and the substrate needed modes, and the modes needed their own state management. Within a few weeks I was running a FastAPI app, a SQLite store with a hundred-plus tables, a sandboxed Python executor, an image generation pipeline, a voice loop, a knowledge pack format, a federated peer protocol, and a 3D avatar substrate. None of which I planned at the start.

Most of it grew because I needed it, daily. I write, so I needed long-form memory that didn't forget. I code, so I needed a workspace that actually understood the project. I read, so I needed offline knowledge packs. I cook, so I needed voice that worked when my hands were full. Augmentum is the shape of my actual life, expressed in software, and that's the only reason it's as broad as it is in this amount of time.

What's built today

Today Augmentum has routed modes for chat, analysis, narrative, agentic deliverables, coding, and a companion surface that remains beta and off by default. It has multi-tenant auth, a bundled inference engine, a sandboxed code executor, local image and voice paths, knowledge packs with hybrid retrieval, an MCP server, and opt-in Fabric peering across machines you own. The substrate is real.

It is also a long way from finished. The install is still Docker Compose, though setup.sh and setup.bat now handle the happy path. The docs are substantial and still catching up. Some surfaces are polished and some are functional-only. That is the nature of building broad: every depth I push reveals four more depths I have not.

What it isn't

It isn't a competitor to the big cloud assistants on raw capability. Their models are bigger. Their RLHF is more polished. Their integrations are deeper into enterprise tooling. If you need the absolute strongest model on the absolute hardest task, you'll probably reach for a cloud provider, and Augmentum will happily forward to one for you when you do.

It isn't a turnkey product for non-technical users yet. The install is Docker Compose; the configuration assumes you know what a port is. I'm working on that. The roadmap converges toward "your parents could install it," but we're not there yet.

It isn't going to be everything. There are categories of feature I won't build because they're not aligned with the project's purpose — anything that requires routing user data through a service I run, anything that creates a vendor lock-in I'd hate from the other side. The orientation is the constraint.

Three things I won't do

These aren't aspirations. If any of them change, I've lost the plot and you should stop supporting the project.

I won't sell you out. No ads. No telemetry without consent. No selling user data — there isn't user data on a server I control to sell, and that's on purpose.

I won't take features away. If something works today, it'll work tomorrow on the same terms. No paywalls bolted onto things that used to be free.

I won't sell out to investors. No funding with strings attached to the two promises above. If Augmentum grows beyond what I can sustain alone, it grows through community or stays small.

Where this goes

I don't have a five-year roadmap. The shape of the next year is clearer: more polish, more documentation, better install for non-Docker folks, mobile companions for iOS and Android, more native UIs for the surfaces that are still functional-only, and a growing community directory of shareable configurations.

Beyond that, the principle is the same: build what I'd want for myself, then make sure other people can use it too. If I'm doing the job right, you'll find Augmentum is the AI you reach for not because it's the most capable, but because it's the only one that lives on your side of the table.

How to be part of it

Use it as your daily AI. One bug from a daily user is worth ten from a curious tire-kicker. Share it with someone it might help — someone who's been priced out of cloud AI, someone with privacy concerns, someone who's been burned by a service shutdown. Contribute back — code, docs, character cards, knowledge packs, translations, anything. Or support development directly if you have the means and want to. No supporter-only features, no badges, no tiered perks. The deal is the same whether you give nothing or five hundred dollars.

-- Matt