I Didn't Need a Reference Manager. I Needed a Thinking System.

• Tags: notion ai paper-management workflow research

I have a confession. I’ve installed Mendeley and Zotero. Each time, I used it for about a week, dutifully importing PDFs and filling in metadata, and then quietly stopped. The library sat there, half-organized, slowly becoming an archeological record of papers I told myself I would read.

The problem was never the tool. The problem was that the act of saving cost more energy than the act of reading.

Recently I switched to Notion + AI, and for the first time in years, the system actually works. This is a short note on why.

The Friction Was the Whole Problem

Reference managers are built around an assumption: that saving is the easy part, and the value comes later — in citation export, library sharing, PDF annotation sync. That assumption probably holds for full-time researchers writing papers every month.

It did not hold for me. What I actually wanted was:

  • Save an interesting paper quickly
  • Understand what it’s about
  • Figure out how it connects to my own work
  • Find it again later when I need it

The traditional tools made step 1 expensive. Download the PDF, fix the metadata, choose a folder, add tags. By the time I got to step 2, my motivation was gone. The library grew, but the reading didn’t.

This is the trap. A reference manager that makes saving feel like work guarantees that most of your library will be unread.

The New Loop

The current workflow is almost embarrassingly simple:

  1. I find a paper.
  2. I drop the link to the AI and ask: “What is this about? How does it relate to my work?”
  3. We discuss.
  4. I say: “Save it to Notion.”

That’s it.

The AI helps me inspect the paper, extract the basic metadata, discuss the main idea, and identify the most relevant place in my Notion archive. Then it creates a new row and writes a structured page inside — Summary, Key Idea, Interesting Points, Discussion, Personal Note. I don’t manually fill in fields. I don’t choose tags. I don’t decide a folder.

What I do is think about the paper. Which is the part that was supposed to matter all along.

What a Page Actually Captures

Here’s the part that took me a while to appreciate. The structured page is not just a summary. It’s a contextualized note.

A paper on neural fields doesn’t just become “NeRF / 2020 / arXiv” in the database. It becomes something like: “useful for thinking about whether my OCT reconstruction model is absorbing aberration into the object representation.” That sentence wouldn’t appear in any abstract. It only exists because the AI and I talked through the paper in the context of my own work.

This is the difference between metadata and thinking. Mendeley stored the former. The new system captures the latter.

Friction Removed, Behavior Changed

When the saving cost dropped to near zero, my reading habits changed.

Papers I would have bookmarked and forgotten now get a five-minute conversation before being archived. Papers I would have skimmed in a browser tab now leave a trace — a paragraph of summary, a couple of open questions, a note about how they connect to something else I read last month.

The Notion guideline I wrote for myself has six reading states: To Read, Skimmed, Reading, Read, Discussed, Revisit. In Mendeley, “Skimmed” was meaningless — the only state that existed was “in the library.” In the new system, even a skim produces real content: what the paper claims, what I understood, what I didn’t. When I come back to a paper three months later, I can see what past-me actually got out of it.

That’s the shift. Status starts to mean something when the system captures thought, not just metadata.

What About Citations?

This is the standard objection. Reference managers exist for citation export. BibTeX, LaTeX integration, citation styles. Surely you still need Zotero for that?

It turns out you don’t. Any paper with a DOI is registered in CrossRef, which exposes a standard content-negotiation API. A single HTTP request, with the right Accept header, returns a clean BibTeX entry:

$ curl -LH "Accept: application/x-bibtex" "https://doi.org/10.1038/s41586-021-03819-2"

@article{Jumper_2021,
  title={Highly accurate protein structure prediction with AlphaFold},
  volume={596},
  ISSN={1476-4687},
  url={http://dx.doi.org/10.1038/s41586-021-03819-2},
  DOI={10.1038/s41586-021-03819-2},
  number={7873},
  journal={Nature},
  publisher={Springer Science and Business Media LLC},
  author={Jumper, John and Evans, Richard and Pritzel, Alexander and
          Green, Tim and Figurnov, Michael and Ronneberger, Olaf and ...},
  year={2021},
  month=July,
  pages={583–589}
}

That’s the AlphaFold paper. All thirty-something authors, full Nature metadata, fetched in well under a second. For arXiv preprints, arXiv has its own BibTeX endpoint. Between CrossRef and arXiv, almost every paper I’ve ever cited is one API call away from a clean entry.

So building a references.bib from my Notion library is not a product feature anymore — it’s a small script. Citation export is not a reason to maintain an entire workflow I don’t otherwise use.

Honest Limitations

I want to be careful not to oversell this. The new system has real gaps:

  • PDF annotation lives elsewhere. If you highlight and margin-note your PDFs, you still need a PDF tool. Notion + AI doesn’t replace that.
  • Group libraries aren’t part of the workflow. Lab-shared collections are still a Zotero strength.
  • AI summaries need verification. The structured page is a draft of understanding, not the understanding itself. I still have to actually read the paper to trust the notes.
  • Cloud dependency is real. If Notion goes down or the connector breaks, the system stops. Zotero stores everything on your machine.

These are not small caveats. For people whose work depends on those features, the trade is probably not worth it.

What Actually Changed

Looking back, I think I had been blaming myself for years for not maintaining a clean library. I just need to be more disciplined. I just need to set aside time every week. The tools were fine; I was the problem.

That was wrong. The tools were optimized for storage, but what I needed was something optimized for reading and thinking. The mismatch was structural, not motivational.

AI didn’t replace the reference manager so much as make the storage layer almost invisible — and once it became invisible, I could finally see what I actually wanted underneath: a place to keep the traces of what I read and thought, not a database of metadata.

The system is light. The point of a paper page is what I thought about it, not what fields I filled in. If maintenance ever starts feeling heavy again, the rule is to simplify, not to push through.

That’s the whole post. The takeaway is small, but it’s the kind of small that changes a habit you’ve been failing at for years.