The thing every reading tool gets right, and the thing none of them do
I read constantly, across articles, books, and podcasts, and for years I retained almost none of it. The tools I reached for were not the problem in any obvious way; Pocket, Instapaper, Readwise, a stack of NotebookLM projects, they all did exactly what they promised. The problem was what they promised. Every one of them is optimized for saving content, and a smaller, better set of them adds a layer of recall on top of the saving, but almost none of them touch the two things I actually needed, which are synthesis and comprehension. Nothing I used could catch a summary that was plausible and wrong, connect an idea in one source to a contradiction in another, or tell me that I had understood a piece rather than merely filed it.
That gap is easy to state and strangely hard to find named anywhere, so I want to be precise about it. Saving is capture. Recall is remembering the thing you captured. Comprehension is a third thing, and it is the one that compounds, because it is the difference between owning a beautiful archive and actually knowing what is inside it.
It is fair to push back here and say that this is a discipline problem rather than a tooling problem, and that no app is going to make me understand what I was too lazy to sit with. I think that critique is half right. But it misreads where the friction actually sits. The reason I never closed the loop was not that the reflection was hard, it was that every tool I owned quietly rewarded the wrong action: the save was one tap and the understanding was unscheduled homework, so the save always won. Tools do not just store your behavior, they price it, and the price was set against the exact thing I cared about.
The reframe that unlocked it
What let me build the thing was not a product idea, it was a pair of readings that collided in the same week.
The first was Andrej Karpathy writing about LLMs as compression engines, the notion that a model distills the statistical structure of human knowledge into something you can query directly in natural language. The second was a critique of LLM "wikis" for organizations, which argued that letting a model rewrite knowledge into itself is dangerous, because a single misinterpretation can quietly corrupt the whole base over time, and that retrieval against fixed sources is the safer design.
Held together, those two gave me the shape of an answer. What would it look like to apply compression to one person's reading life, but keep the sources as the source of truth and only ever retrieve against them? Not a wiki that rewrites what I read into a lossy summary I then trust blindly, but a system where I write a one-line note on what each source meant to me, embed the whole thing, and query my own corpus annotated with my own synthesis. Writing that curator note is itself lossy compression, but with unusually high signal preservation, because I am not archiving the web, I am archiving what the web means to me while the web itself still stands behind it as the record.
What I built, and what it taught me
I built a full-stack app where I save URLs, PDFs, and videos, write what I took from each, and index the whole corpus as vector embeddings. It holds a few dozen sources today, chunked into roughly a thousand pieces, with review questions generated at ingest so the archive can quiz me on what I saved rather than only retrieve it. It runs a Socratic mode that asks me a question before it answers mine, and one session in five hands me a summary that is deliberately, plausibly wrong and waits to see whether I catch it. I use it every morning.
Two things surprised me, and both were the opposite of what I expected walking in.
The first is that curation is far harder than retrieval. Getting the RAG pipeline working took a weekend. Getting myself to write a real note instead of dropping a bare URL is the actual product problem, and it never fully goes away, because the feature that makes the system valuable is also the one that costs the user the most effort. That is an uncomfortable thing to design around, and refusing to admit it is how most of these tools quietly become prettier bookmark folders.
The second is the one I did not see until I had months of my own usage to look back on: the most valuable signal in a reading tool is not what you saved, it is what you abandoned. Every app tracks the former and none track the latter. So the archive now runs a weekly pass that flags the topic clusters I have quietly stopped engaging with and asks me one question about each, which is whether the drift was intentional. Sometimes it was. Often it was a thread I meant to keep pulling and simply lost, and no tool I have ever used would have surfaced that, because none of them were built to notice an absence.
The honest part
I want to end on the thing I have not resolved, because it is the real risk and the essay would be dishonest without it.
There is a narrow slice of people who would rather be challenged than served, and I am plainly one of them, but I do not yet know how wide that slice actually is. Some of the people who think they want a tool that tests them will churn the moment the novelty wears off and the friction starts feeling like friction rather than virtue. So the open question is not whether the gap in the market is real, because it clearly is. The question is whether it stays empty because it is genuinely hard to fill, or because the market for filling it is too small to matter. I am building to find out, and the tool I open every morning is the experiment.