In 2004, two product teams placed the same bet, twelve weeks apart, and neither knew the other was doing it. On April 1, Gmail launched with labels instead of folders, built by Paul Buchheit. On June 28 at WWDC, Steve Jobs walked on stage and demoed Spotlight. Both teams made the same call: people want to find things, not to file them. Twenty-two years later, the call has aged well in some places and badly in others, and most productivity advice still gets the split wrong.
The personal archive has gone search-first. The work-in-flight layer and the shared team layer have not. This post explains the distinction, because conflating them is how good knowledge-management advice turns into bad knowledge-management practice.
Quick takeaways (as of May 2026): Gmail launched April 1, 2004 with the "labels not folders" thesis; Spotlight debuted at WWDC June 28, 2004. Multi-modal vector embeddings (CLIP 2021, OpenAI text-embedding-3 January 2024, Voyage multimodal-3 November 2024) finally made personal-scale search work across photos, screenshots, and PDFs. Bergman and Whittaker's 2016 MIT Press book showed folder preference is a cognitive bias, not a productivity advantage. McKinsey: knowledge workers still spend 19% of the workweek looking for information.
The day filing stopped winning
Buchheit started Gmail in 2001. By the time it launched on April 1, 2004 (people thought it was a prank), it had a feature set that read like a deliberate refutation of Outlook: 1GB of storage, threaded conversation view, integrated search, and labels. The labels choice was the load-bearing one. A message could carry many labels. It did not have to live in one place. The search box was first-class.
Twelve weeks later, Spotlight shipped. The framing in Jobs's WWDC keynote was the same one Buchheit had used: stop deciding where to put a file before you know what you will use it for. The index will find it.
Both products won the personal-archive layer for the next decade. Neither won the active-work layer or the team-collaboration layer, and that nuance is the whole post.
Why your folder tree was always lying to you
George Miller's 1956 paper on the magical number seven, plus or minus two, set the public framing for human short-term memory limits. Nelson Cowan revised the ceiling downward to roughly four chunks under focused attention in his 2001 work. Either way, the cognitive budget for top-level folders is small and you blow through it inside a month of real work.
The orphan problem is the other tax. A meeting note belongs in /clients, /q2-roadmap, and /1on1s at the same time. Pick one and you lose two retrieval paths. Tag all three and you have rebuilt the labels-not-folders thesis poorly. Most people pick one and forget. The note then becomes findable only by search, which means the folder you spent time choosing did nothing.
Ofer Bergman and Steve Whittaker, in The Science of Managing Our Digital Stuff (MIT Press, 2016), surveyed decades of personal-information-management research and arrived at an uncomfortable finding: people prefer folders to search even when search is empirically faster. Their earlier 2008 ACM TOIS work, and a 2015 Nature Scientific Reports paper, traced the preference to the same spatial brain structures we use for real-world wayfinding. Folders feel good because they activate the hippocampus. Feeling good is not the same as being fast.
The contrarian split: search-first for the personal archive, structured for the active and shared layers. Most essays pick a side. Both sides are wrong on their own.
The 19% tax (and why AI just refinanced it)
The McKinsey Global Institute's 2012 Social Economy report put knowledge workers at 19% of the workweek spent searching and gathering information, roughly 9.3 hours a week. A 2024 St. Louis Fed analysis found generative-AI users recovered about 5.4% of work hours, around 2.2 hours a week. The Microsoft and LinkedIn 2024 Work Trend Index reported 75% of knowledge workers were using AI at work, half of them within the prior six months.
Read those numbers together and the picture is clear. The retrieval tax is real. AI is not eliminating it. AI is replacing one tax (filing) with a smaller one (asking the right question). Trading folder discipline for search discipline only saves time if the search actually works on the content you have. That is what changed in 2024.
What changed between 2021 and 2026
The reason search-first is finally credible for the personal archive comes down to four technical shifts in five years.
- CLIP (OpenAI, February 2021). The first model that put text and images in the same vector space. Made it possible to search photos by description without having tagged them first.
- OpenAI text-embedding-3-large (January 25, 2024). 3,072 dimensions, $0.00013 per 1k tokens, 8,191-token context. The cost crash that made personal-scale vector search affordable.
- Voyage multimodal-3 (November 12, 2024). Built for interleaved text and images plus screenshots of PDFs and slides. +19.63% retrieval accuracy over the prior state of the art. Voyage was acquired by MongoDB for $220M in February 2025.
- vec2vec (Jha, Zhang, Shmatikov, Morris, arXiv 2505.12540, May 2025). Unsupervised translation between embedding spaces with cosine similarity up to 0.92 to ground truth. Evidence that embedding geometry is approximately universal, which means your archive is portable across the models that come after the one you used to index it.
Add 1-million-token context windows and voice as a first-class input, and the personal-archive use case is finally a solved engineering problem.
Knowledge is a lake, not a tree
James Dixon, then CTO of Pentaho, coined "data lake" in October 2010. His framing: "If you think of a Data Mart as a store of bottled water (cleansed and packaged and structured for easy consumption) the Data Lake is a large body of water in a more natural state." In data engineering the lake won because pre-deciding schema made you wrong about future questions.
Personal knowledge has the same shape. You cannot pre-file what your future self will need. The folder is a tree. Your memory is a lake. Trees demand that you decide where every leaf goes before it grows. Lakes hold whatever you pour in and let you fish later. Kevin Moody, Mem's co-founder, put it directly when he wrote that no matter how hard we try to impose the perfect folder structure and tagging system, we cannot perfectly anticipate how to organize information for the moment we will actually need it.
Four things to stop filing today
Screenshots
The average smartphone holds around 2,795 photos. US users take about 20 a day. Globally, more than two trillion photos were taken in 2025, 94% of them on smartphones (Photutorial). The folder structure inside your camera roll is theatre. Modern OCR plus visual-language models can read the text inside any screenshot. Stop sorting; start searching.
Merlin Mann's original 2007 Inbox Zero talk at Google was not about a zero counter. He spent most of it on five verbs (delete, delegate, respond, defer, do) and on the idea that the inbox is a place where open decisions wait to be closed. The cult that grew up around the literal number misread him. Cal Newport, in Slow Productivity (2024), put the case bluntly: "No one ever changed the world, created a new industry, or amassed a fortune due to their fast email response time." Archive everything, search when you need it. Gmail had this right in 2004.
Photos
Apple Memories and Google Photos already detect people, places, scenes, and text inside images. You can search "beach, 2022, family" or "the receipt photographed in Lisbon" and get the right shot back. The manually built album is largely decorative.
Receipts and personal documents
Vision models extract amount, vendor, date, and category from any clear receipt in seconds. Tax prep gets dramatically easier when the query is "all medical expenses 2024" rather than "open the medical folder, hope nothing got misfiled". The folder was never the source of truth. The receipt was.
Three things still worth structuring
The honest counterargument, and the part most search-first essays skip.
- Public-facing writing. You want to know exactly where your essays and posts live. Canonical URLs matter for SEO, for citation, and for your own ability to point people at them. Structured layout, every time.
- Active project files. The work in flight needs a stable shape so collaborators (and a future-you walking back into the project after a vacation) can pick it up cold. The right answer is a small, stable directory tree, not search.
- Shared team knowledge. Collaboration requires schema. "Where do new joiners read about deployment?" needs a deterministic answer, not a vector-search result that varies by query phrasing. Notion, Confluence, and Google Docs win this layer for a reason.
| Layer | Search-first lake | Structured tree |
|---|---|---|
| Personal photos and screenshots | Yes | No |
| Receipts, voice notes, forwarded emails | Yes | No |
| Active project files in flight | No | Yes |
| Shared team wiki and runbooks | No | Yes |
| Public-facing essays and posts | No | Yes |
| Archived old projects | Yes | No |
The principle is one sentence: search-first for the personal archive, structured for the active and shared layers. Conflating them produces both bad archives (tidy folders nobody can search) and bad collaboration (free-for-all wikis nobody can navigate).
The retrieval skills replacing filing skills
If filing was the discipline of the previous era, retrieval is the discipline of this one. The skills are different.
- Query phrasing. Specificity, time-bound qualifiers ("last March" beats "recently"), modality hints ("the screenshot with the gate number").
- Natural-language recall. Asking by what something is, not where it is. "The thing a colleague said about retention" beats "the notes folder from Q1".
- Trust calibration. Knowing when to verify AI's recall against the original source, and when the recall is good enough on its own.
- Capture discipline. The lake is only useful if the right water gets into it. Forward, photograph, and voice-note the things you will want later. Stop deciding what to file at capture time.
What this means for the next ten years
When memory is queryable, the bottleneck moves from "capture and file" to "ask and verify". Filing was a tax we paid to make future retrieval possible. That tax is increasingly avoidable. The personal archive of 2030 is going to look much more like a queryable lake of voice notes, photos, and forwarded emails than a manicured folder tree.
MemX is built on this thesis. Forward anything to your WhatsApp memory. Photo of a receipt, voice note after a doctor's appointment, screenshot of a flight confirmation, forwarded email from your contractor. No folder. No tag. No daily-notes ritual. Six months from now you ask "what did the contractor quote for the dishwasher?" and the answer comes back in two seconds because the receipt was read at capture time. The lake stays. The model on top of it can change.
Try the lake model on your own archive: send your next 10 captures (photo, voice, forward, screenshot) to MemX on WhatsApp, then ask for one of them back in plain English a week later. memx.app is free to start.
Key takeaway: folders did not die for everything. They died for the personal archive, where pre-deciding schema is always wrong, and they survived for the active and shared layers, where structure is load-bearing. Get the distinction right and you stop paying the 19% tax that has not budged in fifteen years.
01Are folders really obsolete in 2026?
For the personal archive, mostly yes. For active project files and shared team knowledge, no. Conflating those two layers is the most common mistake in productivity advice. The personal archive belongs in a queryable lake; collaboration belongs in a small, stable structured tree.
02What killed folders for the personal archive?
Three shifts: multi-modal embeddings (CLIP onwards) putting photos, screenshots, and text in one searchable space; OpenAI's January 2024 embedding price crash making personal-scale vector search affordable; and voice as a credible input format. Together they made retrieval cheaper than filing.
03But research says people prefer folders. Is search really better?
Bergman and Whittaker (MIT Press, 2016) found preference does not match performance. People navigate folders even when search is faster because folders activate the same spatial brain structures as physical wayfinding. The bias is real. Performance still favours search for the personal archive.
04What about email? Should I delete my folder structure?
Gmail had this right in 2004. Archive everything, search when you need it. Merlin Mann's original Inbox Zero was about closing open decisions, not maintaining a zero counter. The folder structure most people built on top of email is theatre.
05If folders die, what skill replaces them?
Retrieval literacy: query phrasing with specificity and time bounds, natural-language recall (asking by what something is, not where it lives), trust calibration on when to verify AI's answer against the source, and the discipline of capturing into one searchable place instead of many manicured ones.
06Does this mean I should delete all my folders?
No. Keep folders for active projects, shared team docs, and anything you ship publicly. Let the personal archive (screenshots, photos, receipts, voice notes, forwarded messages) flow into a single searchable place. The mistake is using the same tool for both jobs.
