Building ASMOS: My Personal AI Operating System

For years my digital life looked like everyone else’s: a folder called code with dozens of repos inside, a notes app here, a chat thread there, and an agent in Cursor that had amnesia every time I opened a different window.

That worked until it didn’t. I wanted one place where my notes, my agent memory, my skills, and my projects could see each other — without turning everything into a monorepo.

So I built ASMOS.

What ASMOS is

ASM (Asbjørn) + OS (operating system) = ASMOS.

One folder on disk. One Cursor window. Agents get the whole picture.

It is not a git monorepo. It is a workspace OS:

Layer	Role
AGENTS.md	The manual — schema, workflows, agent rules
wiki/	A Karpathy-style personal knowledge wiki
wiki/memories/	Long-term agent memory with decay and forgetting
.cursor/skills/	Shared capabilities (ingest, remember, curate, …)
repos/<name>/	Nested code projects — each keeps its own git remote
capture/	Inbox for stuff waiting to be ingested into the wiki

The outer repo tracks the OS itself. Code inside repos/ pushes to GitHub independently, the way it always did — except now agents can read across repos without me juggling windows.

The problem I was solving

Most “AI setups” are really collections of freelancers:

• ChatGPT doesn’t know what Cursor did yesterday.
• Cursor doesn’t know what’s in my notes.
• Every repo is an island with its own rules, if any.

A real team shares context, vocabulary, history, and goals. I wanted infrastructure for that — a persistent layer above whichever model or vendor I rent this week.

Own your context. Rent your intelligence.

ASMOS holds the memory. Models are interchangeable collaborators.

Design influences

I stole shamelessly from people smarter than me:

• Karpathy’s LLM Wiki — immutable raw/ sources, distilled wiki/ concepts
• Hermes Agent memory — bounded hot cache (MEMORY.md, USER.md), importance decay, archive instead of delete
• repo-of-repos — read everywhere, write in one repo at a time
• Nate Herk’s AIS-OS — one root folder, “other worlds” under repos/

The result feels less like a notes app and more like an operating system for agents.

How I actually use it

Capture → ingest

I dump quick notes, chat exports, and fragments into capture/. No frontmatter, no ceremony — presence in the folder means pending.

When I’m ready to process the inbox (and I treat that as a ritual, not a chore), I open ASMOS in Cursor and run @ingest. The agent:

1. Picks a home under raw/<subfolder>/
2. Writes or updates wiki concepts with links to the archived source
3. Moves the file out of capture/ and stamps it as ingested

Empty inbox. Knowledge filed. I stay in the loop without babysitting YAML.

Memory that forgets on purpose

Agents write atomic memories to wiki/memories/active/ with importance scores. A scheduled script applies cognitive-style decay — old unused facts fade and eventually archive. Frequently accessed memories stay warm.

No review queue. No “please confirm you still want to remember this.” The system drifts toward what I actually use.

Nested repos, one window

I’m migrating projects from ~/Documents/code/ into repos/ in batches. This blog lived in a weird nested folder (asbjborg_site/asbjborg.github.io) until today — now it’s repos/asbjborg.github.io/ inside ASMOS, still pushing to the same GitHub remote.

Still early

ASMOS is a weekend’s worth of architecture and a week’s worth of migration — not a product launch. The wiki is sparse. Most of ~/Documents/code/ is still outside the OS.

But the loop works: capture, ingest, remember, code — all in one tree. Codex, Cursor, and ChatGPT can all read the same USER.md and AGENTS.md and actually know who they’re talking to.

If you’re building something similar, the portable part is the architecture above — ASMOS itself stays a private personal repo for now, not an open-source release.

More posts as the migration continues. For now — this one is proof the nested-repo setup survives a real push. 🚀