AI File Manager Put to the Test: How Marvis Tackles a Nearly Full 1TB Drive

When AI Can Write Proposals and Do Research, Why Does Nobody Use It to Organize Their Computer?

I use AI every day to draft proposals, conduct research, and organize content. I've handed off all kinds of complex tasks to AI — but when it comes to something as basic as organizing my computer, almost nobody takes it seriously. Open my Downloads folder and you can scroll all the way back to files from 2022. Years of accumulation have turned it into complete chaos.

To be fair, various AI agents have been capable of controlling and organizing local files for the past couple of years. But I've always only dared to let AI tidy up my desktop — nowhere else. My biggest worry: AI might enthusiastically move everything, re-categorize it all, and build a bunch of folders that look neat but leave me unable to find a single thing. Like one of those hyper-organized Japanese-style moves — everything is tidy, but you're standing there completely lost.

This time, I tried out Marvis's AI file management feature to see if it could help me do a proper "cyber declutter."

Organizing Without Moving Files: The "Index First" Approach

What I cared about most was how it actually organizes files. My old mental model of organizing was simple: move things around. Take the PDFs in Downloads and move them to "Study"; move images to "Life"; move Word docs to "Work." Sounds reasonable — but for a computer that's already a mess, with files piling up in waves over the years, the last thing I want is to go through all that again.

Marvis's approach surprised me. Instead of moving anything, it first recognizes your files within its own interface — like building a directory layer on top of your computer. It automatically sorts everything into complete categories: certificates, contracts, data, receipts, books, reports, and more. The key part: it doesn't touch the original files. Click on any file and it tells you exactly where that file lives on your actual drive — whether it's in Downloads or on the Desktop — without changing a single thing.

It's more like hiring a very sharp-eyed warehouse manager who walks through the entire storeroom, reads every box, sticks on labels, and draws a map — while everything stays exactly where it is, ready for you to move or delete at your own discretion. For someone like me who's terrified of AI making unauthorized moves, this is huge.

This approach is technically known as index-based management, as opposed to migration-based management. Traditional file organization tools — including early file management scripts and Automator rules — mostly work by physically moving files, which is hard to undo once executed and extremely costly if something goes wrong. An index-based solution maintains a local metadata database that records each file's path, type, content summary, and custom tags. The presentation layer and the storage layer are completely decoupled. What you see is a "tidy view"; the actual directory structure on disk remains untouched. macOS Spotlight and iOS Photos smart albums work on the same principle — search and categorization results are just mappings to the original files, not reorganizations of them. This is especially safe for file libraries built up over many years, since many apps, scripts, or workflows may depend on absolute file paths — a physical move would break those references instantly.

Following the Thread: Connecting Scattered Pieces of One Project

Another detail I found genuinely useful: click on a file — say, a trip itinerary summary — then hit "Expand," and it pulls up everything else left behind by the same project: contracts, product catalogs, spreadsheets, Word documents, images. The whole chain surfaces at once.

Before, searching for files meant hunting them down one by one, with the rest left to luck. Now, opening one file brings the entire thread along with it — no more piecing things together yourself. Files that used to be scattered in different corners, completely isolated from each other, can now appear together as a single, complete "event."

Finding Files by Fuzzy Memory, Not Exact Filenames

Every adult who's used a computer for years knows this painful moment: you know the file is somewhere on your drive, but the second you ask yourself "what was it called?" — your mind goes completely blank.

I had a very real use case: I wanted to find all the materials related to my travel group projects for my startup, but I had absolutely no memory of what I'd named any of them. So I just opened a new chat and asked it to find "itineraries, quotes, and client requirement docs I'd done for CET."

It went through the locally indexed files and listed everything that might be relevant. Clicking through, I could actually see the itineraries and quotes from back then — and some materials I had completely forgotten I'd ever saved. This feature is perfect for anyone with a messy computer: you don't need to remember the exact filename. A rough impression is enough.

The core technology here is semantic search, which is fundamentally different from traditional keyword-based full-text search. Traditional search requires the query to match strings in filenames or file content exactly. Semantic search, by contrast, uses an embedding model to convert both the file content and the user's query into high-dimensional vectors, then calculates the cosine similarity between those vectors to determine semantic relevance. This means that even if the words you type are completely different from the words in the file, the system can still surface it — as long as the meaning is similar. For example, searching "trip plan" might match a file containing "route planning" or "itinerary." Implementing local semantic search typically requires either running a lightweight embedding model on-device or sending file content summaries to a cloud API — which is why privacy policy transparency is a critical concern for this category of AI file tools.

Batch Renaming: Preview First, Execute Later

Finally, I tested a feature I really needed: batch file renaming. I had a bunch of photos taken in Guizhou, all named with some variation of IMG followed by a string of numbers, spanning a long stretch of time. I had no idea which photo was which, let alone how to pick footage for a video edit.

Marvis first generates a comparison table showing the old and new filenames. Only after I confirm everything looks right does it execute. Hundreds of photos renamed, fully under my control the whole time. I especially appreciate the "preview first, execute later" step, because the last thing I want is AI making unilateral decisions — especially when hundreds of photos are involved. That confirmation step is absolutely necessary.

This "preview, confirm, then execute" interaction pattern is known in human-computer collaboration as Human-in-the-Loop. The design philosophy: AI generates the plan, humans make the final call, and execution authority always stays with the user. For irreversible bulk operations like mass file renaming, this mechanism is critical — if hundreds of files are incorrectly renamed, photo apps, video editing project files, or backup indexes that rely on the original filenames could all break simultaneously. The comparison table (a diff preview) lets users verify at zero cost whether the AI's naming logic matches their expectations — did it correctly identify the shooting date, location, or event name? — before deciding whether to execute fully, partially, or regenerate with adjusted prompts. This shares the same "changes visible, execution controlled" design philosophy as Git's git diff-before-commit workflow and the Pull Request mechanism in code review.

Organizing Your Computer Is Also a Cyber Walk Down Memory Lane

I've had this main machine for four years. It's seen me through a full-time job, a resignation, graduate school, a first startup, and a second startup. Digging through the files felt like watching a cyber documentary of my own life.

I found my application materials for a journalism master's program in Hong Kong, and screenshots of IELTS notes from when I was still studying for the test. That whole stretch — working a day job while applying to programs, anxious about the future every single day — all crammed into one folder. The filenames were a mess, which means I was a mess back then too. But those files really did carry me through that period, one by one. I used to think of them as clutter taking up space. Looking at them again this time felt strange in the best way — it turns out an older version of me has been living in this computer all along.

Honestly, there was no "one click to save the 1TB drive" miracle here. But at least now I know which files are junk, which are duplicates, which are in the wrong place — and I can ask Marvis in the chat window to help me delete them or create new folders to sort things out.

I used to think organizing files was a minor task, something easy to push off until later. What I've come to understand now is this: no matter how much you use AI, the real productivity tool is the clean, organized computer underneath it all. Spending a little time finding it a good cyber steward is, in its own way, a small act of care for your past self.