Your AI has amnesia. I got tired of paying for it.
Here's a small, stupid moment that changed what I've been building for the last stretch of my life.
I was three hours into a hard problem with an AI coding assistant. We'd finally worked out the shape of the thing — the tradeoffs, the dead ends I'd already walked down, the one weird constraint that made the obvious solution wrong. It was good. It felt like working with someone sharp.
Then the session ended. And the next morning, it knew none of it. Not the constraint. Not the dead ends. Not the decision we'd made together and why. I opened a fresh chat and started re-explaining my own project to a tool that had, the day before, understood it better than most people would.
I did that again the day after. And the day after that.
If you've used any of these tools for real work, you know exactly the feeling. The models are genuinely brilliant — and they have the memory of a goldfish. Every session starts from zero. You become an unpaid narrator, re-reading your own life to a stranger who's very clever and completely new every single morning.
That's the problem I set out to fix. The tool I built to fix it is called Anvil, and this is the first of a weekly series where I'll walk through it one piece at a time — what each part does, why it exists, and how it helps. Consider this the introduction.
The problem isn't intelligence. It's memory.
We've spent two years making AI smarter. Almost nobody has made it remember.
Think about what that actually costs you. Every time you re-paste the context, you're paying — in time, in tokens, in focus. Worse, the assistant repeats mistakes it "learned" not to make yesterday, because yesterday is gone. It has no idea it already tried that approach and it already failed. You're the only one carrying the thread, and you're carrying it in your head and your scrollback.
The tools that do offer "memory" mostly bolt on a little cloud key-value store: a handful of facts about you, living on someone else's server, shallow and per-account. Better than nothing. But it's not a brain. It's a sticky note.
I didn't want a sticky note. I wanted the thing to actually know my work — and I wanted that knowledge to be mine, on my machine, not rented back to me from a cloud I don't control.
So what is Anvil?
Anvil is a private, local, structured memory-and-reasoning layer that sits underneath the AI you already use.
It doesn't replace your assistant. It plugs into it. Claude, and others — you keep using the tool you like, signed into their app the way you already do. Anvil quietly becomes the long-term memory and the sharper reasoning behind it, so the assistant you have suddenly remembers, learns, and thinks harder than it can on its own.
Three things make it different from a sticky note, and they're the spine of everything I'll cover in this series:
- It remembers — properly. Not a flat list of facts, but a structured knowledge graph: your projects, the decisions, the relationships between them, with the right pieces recalled automatically on every prompt. It gets denser and more useful the more you work, instead of resetting.
- It learns from its own work. When a mistake happens twice, Anvil can turn that lesson into a hard rule it won't cross again. It distills your accumulated work into reusable skills. The system that helped you last month is measurably better this month — not because a vendor shipped an update, but because your history taught it.
- It's yours. Everything lives on your machine, encrypted. Your memory, your keys, your projects. Anvil does its best to keep your secrets out of whatever gets sent to a model. There's no cloud lock-box holding your knowledge hostage. If your internet drops, if a subscription lapses, if a provider has a bad day — Anvil degrades a rung and keeps working. It never bricks, and it's honest with you when something's off instead of pretending everything's fine.
That last point matters more to me than any feature. A tool you rely on for real work has to be trustworthy about its own limits. I'd rather Anvil tell you "I couldn't verify this" than dress up a guess as a fact.
The part people don't expect
Here's where it stops being "a memory add-on" and starts being something I'm genuinely proud of.
Anvil can convene what I call a Crucible — a small cast of characters, each fused with real expertise, who argue a hard question out loud and resolve it to a sharper answer than any single pass would give you. A skeptic. A first-principles thinker. A devil's advocate. They disagree, they pressure-test, they converge. It's the difference between asking one clever person and running a decision past a room.
And because I think a brain you can't see is a brain you can't trust, Anvil renders its own mind as something you can look at — a real 3D galaxy where every memory, skill and project is a star you can fly through. When it's thinking, you can watch it think: the Crucible's voices orbiting the core, the work lighting up the clusters it touches.
I'll give the Crucible and the galaxy their own articles later in this series, because they each deserve one. For now I just want you to know they exist, and that they're not gimmicks — the galaxy is how I actually debug what the system knows, and the Crucible is how I make it reason about things that matter instead of blurting the first plausible answer.
Who I built this for
I built Anvil for people who do real, sustained work with AI and are tired of the goldfish.
Developers, obviously — that's my world. But the shape of the problem is universal: anyone whose work has continuity, who makes decisions today that should inform tomorrow, who is sick of re-explaining themselves. And I built it especially for people who care where their data lives. If the idea of your working knowledge sitting in a cloud you don't control bothers you — it bothers me too. Anvil's answer is simple: it doesn't.
There's also a quieter reason. I believe these tools should be mentors, not ghostwriters. Anvil is built to make you better at your work, with guardrails that keep the work yours — not to quietly do it for you and hollow out your own skill. That principle runs through the whole thing, and you'll see it show up again and again as this series goes on.
A day with it, versus without it
Let me make this concrete, because "memory layer" is the kind of phrase that means nothing until you feel it.
Without Anvil, a typical morning looks like this. You open your assistant. You paste the same paragraph of context you pasted yesterday — the stack, the constraints, the thing you're mid-way through. You get going. Halfway in, the assistant confidently suggests the exact approach you already ruled out on Tuesday, because Tuesday is gone, so you explain again why it doesn't work. You make a good decision together. You close the laptop. Tomorrow, all of it evaporates, and you do the whole ritual over. The tool never gets to know you. You are permanently on the first date.
With Anvil, you open your assistant and just… start. The context is already there — recalled automatically, the right pieces for this prompt, pulled from a structured memory of everything you've done. The approach you ruled out on Tuesday is a logged decision with the reason attached, so it doesn't get suggested again — and if it somehow does, the rule catches it. The thing you were mid-way through is on a live map of the project. You're not narrating your life to a stranger. You're continuing a conversation with something that has been paying attention the whole time.
The gap between those two mornings is the entire reason Anvil exists. It's not a bigger model. It's continuity — the one thing raw intelligence can't give you on its own.
Who's behind this
I'm building Anvil under Vitkilabs. I'm not a big company with a marketing department; I'm a builder who hit this wall hard enough, often enough, to stop complaining and start soldering. That's worth knowing, because it shapes the product: every decision in Anvil is one I made because I needed it to be that way for my own work, not because a growth chart asked for it.
It also shapes how I'll write these. I'm not going to hype. I'm going to show you the thing, tell you honestly what it does and doesn't do yet, and let you decide. If a piece is half-built, I'll say "half-built." If something's genuinely good, I'll show you it running rather than asking you to take my word.
What this series is going to be
Every week, I'll take one piece of Anvil and go deep on it — plainly, honestly, no jargon-as-marketing:
- What it is — in human terms.
- Why it exists — the actual problem it solves, usually one I hit myself.
- How it helps — what changes for you when you have it.
Roughly where we're headed after this intro: the memory brain itself, the Crucible, the built-in security range, the live galaxy, local-first privacy, the skills-and-rules learning loop, self-optimising personas, how it routes across providers and survives offline, the project-management layer, endless-session distillation, and the own-your-data licensing model.
Where it helps, I'll bring animations — because honestly, some of this is easier to watch than to read, and I'd rather show you the thing moving than describe it at you.
An honest note, because that's the whole point
Anvil is early. I'm building it largely in the open, and I'm not going to pretend it's finished or flawless — that would violate the exact principle the tool is built on. Some pieces are rock-solid and shipping today; some are still being sharpened. I'll be clear about which is which, every time. When it improves, updates flow to you — that's what the update system is for.
Registration is opening soon, on macOS, Windows and Linux. The core memory brain will be free — not a crippled trial, the actual brain, yours to keep. I'd rather earn a paid tier by being genuinely worth it than by holding your own memory for ransom.
If the goldfish thing has been quietly driving you up the wall too — follow along. Next week I'll open up the memory brain itself: how Anvil decides what's worth remembering, how it hands the right piece back at the right moment, and why "structured" beats "a pile of text" every single time.
Thanks for reading. I'm glad you're here for the start of it.