If We Invented a File System Today

If we invented a file system today, what would it look like, and what would we call it?

The file system on my computer is full of folders that made sense at the moment I created them. There are folders for projects, trips, taxes, receipts, photos, and a growing number simply named after a year. Inside those are more folders, along with files named things like final, final-2, and occasionally final-final.

This is not because I don’t care about organization. The problem is that organizing a file requires me to predict how I will look for it in the future.

Should the receipt for a laptop go in Receipts, Taxes, Work, or the folder for the year I bought it? The answer might be all four, but a traditional file system wants one location. I can make copies or aliases, but then I have to remember which is the original and whether the copies are still current. Usually I choose one folder, move on, and hope my future self thinks the same way my present self did.

Folders were a brilliant metaphor for making computers understandable. They mapped something unfamiliar to the physical world of filing cabinets, folders, and documents. But the metaphor also carried over the limitations of that world. A piece of paper can only be in one folder at a time. A file on a computer doesn’t need to have that limitation.

Search has helped, but it has mostly been layered on top of the same old structure. It works well when I remember part of a filename or a phrase inside a document. It works less well when I remember the circumstances around a file: the presentation someone sent me before a meeting, the photo of a receipt from a work trip, or the document where I wrote down an idea related to something I’m working on now.

AI may finally make a different model practical.

Imagine that saving something did not begin with choosing a name and location. The system could understand what the item is, where it came from, when it was created, who was involved, and how it relates to things already on the computer. A receipt could simultaneously belong to the laptop it documents, the trip during which it was purchased, the expense report it supports, and the tax year in which it matters. These would not be four copies. They would be four useful ways of reaching the same thing.

Instead of navigating to a location, I might ask:

Show me everything related to the kitchen remodel that I haven’t sent to the contractor.

Or:

Find the document Sarah sent before our meeting in Seattle.

The result would not just be a list of files containing matching words. It could include messages, photos, calendar events, notes, web pages, and files from different apps because the system understands the relationships among them.

In this model, folders could still exist. Sometimes a manually curated collection is exactly what I want. But folders would become views rather than containers. I could organize the same information by project, person, place, date, or any other context without moving or duplicating anything. The system might even create temporary views in response to a question and let me save the useful ones.

For this to work, the underlying system would need to know more than our current file systems do. Every item would need a durable identity that does not change when it is renamed or reorganized. Relationships and history would need to be first-class data. The system should know where something came from, what changed, and which other item it was derived from. Applications should contribute to this context without owning the only copy of it.

There are obvious reasons to be cautious. A system that understands all of this would also have the potential to understand far more about us than we intend. It should be local and private by default. We would need to see why a result appeared, correct mistaken relationships, and control which applications and people can access which parts of it. AI should help interpret the data, but it shouldn’t become the only way to retrieve it. The basic identity, history, and relationships should remain dependable even when the AI gets something wrong.

That last point seems important. Today’s file systems are limited, but they are also predictable. I know that if I put a document in a folder, it will be there later. Any replacement has to preserve that confidence while making organization less dependent on my ability to anticipate the future.

So what would we call it?

Calling it a file system would undersell the idea. The word file describes the artifact, but the more useful part of this system would be its understanding of context. Names like knowledge graph or semantic store may be technically accurate, but they are not the sort of names most people would want to use.

I think I would call it Memory.

Memory is not a perfect archive. It is associative. A person reminds us of a place, a place reminds us of an event, and an event brings back a conversation. We rarely remember the exact drawer where a piece of information belongs. We remember what it meant and what it was connected to.

That feels like the better metaphor for a computer designed today: not a filing cabinet that asks us where everything should go, but a memory that helps us find our way back.