# Remember What You Read: Stop Hoarding > Dan Koe's project-first method has a useful core, but memory science adds two missing steps: retrieval and spaced reuse. _Source: Dan Koe X Article, checked against cognitive-science reviews · 2026-07-28 · 9 min read · Verified against primary sources_ Canonical: https://iyu.app/e/remember-what-you-read-dan-koe ## The 60-second version Project-first reading becomes more reliable when it includes retrieval, correction and spaced reuse. **Key points** - Choose an output before choosing what to study. - Close the source and reconstruct the idea rather than trusting familiarity. - Use a curated commonplace book to support creation, not consumption tracking. - Let AI retrieve and question material, but keep evidence and authorship under human control. **Verdict.** Koe's workflow is a useful selection and creation system; memory science supplies the mechanism that makes retention less accidental. ## Full explainer Dan Koe's argument is deliberately contrarian: stop trying to retain every sentence. Begin with a meaningful project, seek only the knowledge that moves it forward, and turn what you find into something you create. That is a strong antidote to collecting highlights without using them, but it is not a complete theory of memory. > **✓** Evidence boundary: Koe's article and workflow are directly verified as his own published argument. The scientific support below applies to components such as retrieval, generation, self-explanation and spacing; it does not prove Eden, Obsidian, a second brain, or Koe's complete workflow as a package. ### THE CLAIM — Why a goal changes what you notice Koe describes learning as a feedback system. A goal defines a desired state; action exposes the gap between where you are and where you want to be; that gap tells you what to learn next. Instead of studying an entire video editor, start making one video and search for the exact technique blocking the next shot. > A goal creates the error signal; the error signal creates a filter; the filter creates relevance. This is best read as a method for selecting information, not a guarantee that relevant information will automatically enter long-term memory. Attention and meaning help, but people still forget material they care about. The practical fix is to add an explicit recall-and-feedback loop. ### EVIDENCE — What memory research supports - **Retrieval practice:** Close the source and reconstruct the argument. Meta-analyses generally find better delayed retention than equivalent restudy. - **Generation:** Predict, complete or restate an idea before checking. Producing an answer can beat merely reading it. - **Self-explanation:** Explain why a claim works and how it connects to prior knowledge; vague paraphrase is not enough. - **Spacing:** Retrieve again after a delay. There is no universal day-1/day-3/day-7 schedule. - **Highlighting and rereading:** They can assist review, but are low-utility as stand-alone strategies and can create misleading familiarity. > **!** Do not replace one absolute with another. 'If it matters, you will remember it' is a memorable slogan, not a dependable cognitive rule. Highlighting and rereading are not useless either; they are simply weaker when they remain passive and untested. ### WORKFLOW — A five-step reading loop - **1. Name the output.** Decide what you are building, deciding or explaining before opening the source. - **2. Read for the current bottleneck.** Capture only claims, examples or questions that can change the next action. - **3. Close the source.** Write a blank-page outline, answer questions or explain the idea aloud without looking. - **4. Check and correct.** Compare your reconstruction with the source so confident errors do not harden. - **5. Reuse after a delay.** Apply the idea in a draft, decision, conversation or project, then retrieve it again later. - **1** — concrete output before research - **0** — need to save every sentence - **2×** — minimum encounters: recall now, reuse later ### NOTES — Build a commonplace book, not a warehouse Koe calls the useful version a digital commonplace book or 'second subconscious.' Its job is not to prove how much you consumed. It should preserve selected ideas that shape your point of view and resurface them when a project needs them. Historical commonplace books mattered because their owners wrote, designed, argued and invented with the material. His tool examples range from Claude Code connected to an Obsidian vault to services such as MyMind and his own Eden product. Semantic search can retrieve conceptually related notes even without matching keywords, but automation does not solve bad curation. A smaller collection with clear provenance and frequent reuse is usually easier to trust. ### AI — Use AI to remove friction, not authorship A useful assistant can search a vault, suggest questions, challenge an explanation, organize a braindump or point out missing evidence. It should not silently manufacture the writer's beliefs. Keep source links, inspect summaries, and make the final argument yourself; otherwise the system may produce fluent prose without durable understanding. ### TAKEAWAY — The test is what you can do without the page Before saving another highlight, close the book and state the main claim, its mechanism and one place you will use it. Then check your answer and revisit it after a delay. The goal is not perfect recall of a library. It is reliable access to the few ideas that improve your next piece of work. ## Primary sources - [Dan Koe — How to remember everything you read (stop trying)](https://x.com/thedankoe/status/2081415714636996844) - [Dunlosky et al. (2013) — Improving Students' Learning With Effective Learning Techniques](https://doi.org/10.1177/1529100612453266) - [Rowland (2014) — Testing versus restudy meta-analysis](https://doi.org/10.1037/a0037559) - [Cepeda et al. (2006) — Distributed practice meta-analysis](https://doi.org/10.1037/0033-2909.132.3.354) - [Bisra et al. (2018) — Self-explanation meta-analysis](https://doi.org/10.1007/s10648-018-9434-x) - [Bertsch et al. (2007) — Generation effect meta-analysis](https://doi.org/10.3758/BF03193441) --- _Published by iyu (https://iyu.app) — the day's AI news, checked against primary sources and rewritten in plain language. Free to quote with attribution and a link to the canonical URL._