My Second Brain

For over 2,120 consecutive days I have reviewed my notes. In that time I have tracked 386 books, collected 16,600 highlights, and reviewed 34,224 of them.

This is not a finished system. It is the current state of something I have been rebuilding continuously for six years, and every layer in it exists because the previous version fell short somewhere. Here is how it works, what it actually delivers, and what I am changing next.

Part one: the system

Capture — Readwise. Everything lands in one place: Kindle, Kobo, email newsletters, Instapaper, X, and anything I type in myself. One rule governs this layer — saving costs nothing and commits me to nothing. Friction at the point of capture is what kills these systems, so there is none.

Subscriptions — Readwise Reader. Articles and a subscription list that would otherwise be unmanageable, including more than 170 podcasts.

The filter — Hermes. Video was the hardest input to handle. I used Snipd to pull insights out of it, and recently replaced that with my own agent. Every morning Hermes sweeps my YouTube subscriptions and, for each new video in the categories I care about, writes a summary with the key takeaways and files it into Reader with an explicit verdict: ignore, keep the takeaways, or watch it.

That verdict is the point. The agent's job is not to gather more — it is to protect my attention. Most mornings the correct answer is ignore, and now something other than me says so.

Views — Obsidian, Notion, NotebookLM. One corpus, three ways of interrogating it. Obsidian for the graph, where connections between topics become visible. Notion for readable summaries of the books I have read. NotebookLM when I want to interrogate a body of material conversationally.

Review — daily. Active recall and spaced repetition, built into Readwise. This is where the 2,111 days come from, and it is the only part of the system I have never skipped.

Part two: what it delivers, and what it does not

The honest accounting has entries on both sides.

What works:

• Retrieval on demand. When a topic comes up, I can find what I have read about it across books, podcasts and articles — without remembering where any of it came from.

• Ideas outlive the book. The aha moment while reading has a short half-life. Captured, connected and revisited, it survives long enough to be useful months later.

• Filtering compounds. Every video Hermes marks ignore is time returned, at no cost to me.

• Connections surface on their own. The graph view regularly shows me that two authors I read years apart were making the same argument.

What does not:

• Volume is not retention. Do the arithmetic: 16,600 notes against 34,224 reviews is roughly two passes per note across six years. That is a glance, twice, a long time apart. Re-reading a highlight produces recognition, which feels like knowledge and is not — the fluency is real and the recall is not.

• The corpus has no memory of what proved true. A highlight I saved in 2019 because it sounded clever carries exactly the same weight today as one that has since proved itself a dozen times. Nothing decays, nothing gets promoted.

• Synthesis is still entirely manual. The system is excellent at capture and decent at review, but I am the one who has to notice the pattern. That does not scale with the collection.

• Tool sprawl is a real cost. Five tools means five things that can break, and periodic effort spent keeping them talking to each other.

Which means the fair summary is this: I built something very good at holding things, and only recently started fixing the part that makes them retrievable.

Part three: what I am changing now

Distil and publish. The newest layer is this blog. Each week I take one idea, synthesize it across every source that converged on it, and write it up in my own words. Writing is the step that exposes how well you actually understood something — you find out the moment you try to explain it to someone else.

Flashcards for recall. Each published idea also becomes a small set of Anki cards, which go into a public deck and into my daily review. This is the direct answer to the recognition problem: a question I have to answer without the text in front of me is a fundamentally different act from re-reading a highlight. Fewer items, harder work, and something that is actually there when I reach for it.

Memory palaces. I built an iOS app, Memoria Palace, for method-of-loci practice. The obvious next step is connecting it to the flashcard layer — importing cards as topics and generating images that combine a locus with the term being memorized. Ideas I want permanently, rather than merely retrievably, would live there.

Techniques I am evaluating. Three ideas are converging on the same conclusion from different directions, and I am watching all three. Andrej Karpathy's LLM wiki: let a model compile and maintain the knowledge base, since the bookkeeping that makes every human abandon their personal wiki is exactly what machines are good at. The Infinite Brain methodology from the AI Impact channel: restructure knowledge for machine traversal so queries get dramatically cheaper, with a harness layer that promotes ideas from private into shared canon only once they have survived review. And Google's Open Knowledge Format, published in June 2026 — a vendor-neutral specification for exactly this shape, plain Markdown with YAML frontmatter, one file per concept.

When an individual, an open-source community and a cloud vendor all arrive at the same answer independently, it is worth paying attention.

None of this is finished, which is rather the point. The system I have today is better than the one I had a year ago, mostly because I kept noticing what it failed to do.

Why it matters

For anyone building a serious reading or note-taking practice:

• Capture is the easy part. Most people optimize the input layer because it produces visible progress. Retrieval is where the value is, and it gets far less attention.

• Volume is the wrong metric. Highlights saved and streaks maintained measure input, not retrievability. It is entirely possible to be disciplined and ineffective at the same time.

• The discomfort is diagnostic. If your review routine feels smooth, it probably is not working. Difficulty during recall is the signal that something is being built.

• Own words are not a stylistic preference. Rewriting an idea creates the connections that later act as retrieval cues. Copying verbatim skips that step, which is why quotation-heavy collections stay inert.

• The scarce resource is attention, not storage. The most valuable component in my stack is the one that tells me to skip things.

• Treat it as a system under revision. Every layer here replaced something that was not working. Expecting to get it right at the outset is why most second brains are abandoned.

Sources

• Ultralearning — Scott Young

• How We Learn — Stanislas Dehaene

• How to Take Smart Notes — Sönke Ahrens

• Super Learning — Peter Hollins

• Memory Power 101 — W. R. Klemm

• Andrej Karpathy on the LLM wiki (April 2026 gist, referenced in Google's OKF announcement)

• Open Knowledge Format — Google Cloud, June 2026

• The Ultimate FREE AI Operating System (Infinite Brain) — AI Impact

Flashcards

These are the cards I use for spaced repetition on this material. The whole Second Brain deck is public on AnkiWeb if you want it.

Q1: What single rule should govern the capture layer of a knowledge system?

Saving costs nothing and commits you to nothing — friction at the point of capture is what kills these systems.

Q2: What is the job of an agent that filters incoming content?

To protect attention rather than gather more — the highest-value verdict it can return is usually "ignore".

Q3: Why keep one corpus but several views of it?

Different tools answer different questions of the same material: a graph shows connections, summaries give readable overviews, and a conversational interface allows interrogation.

Q4: Why does re-reading highlights feel effective when it isn't?

It produces fluency and recognition, which are mistaken for knowledge. Familiarity with a corpus is not the same as being able to use any of it.

Q5: What happens when passive review is compared with retrieval practice immediately after study, and days later?

Immediately after, passive review looks better. Days later, retrieval practice beats it by a wide margin — the technique that felt worse is the one that worked.

Q6: Why is recall not simply playback of a stored file?

Remembering is a creative act of reconstruction from fragments, not retrieval of a saved copy — which is why the act of recalling strengthens the memory.

Q7: What is the difference between storage strength and retrieval strength?

Storage strength is how deeply something is embedded and may only grow over a lifetime; retrieval strength is whether you can reach it right now, and it decays. Retrieval strength is the one that matters.

Q8: What structural weakness do most highlight libraries share?

No mechanism for promotion or decay — an idea saved because it sounded clever carries the same weight as one that has proved itself repeatedly.

Q9: Why is rewriting an idea in your own words not merely a stylistic choice?

The act of rewriting builds the connections that later serve as retrieval cues; copying verbatim skips that step, which is why quotation-heavy collections stay inert.

Q10: Name the techniques with the strongest evidence for learning and retention.

Practice testing, distributed practice, elaborative interrogation, self-explanation, and interleaved practice.

Q11: Why is bookkeeping the argument for letting a model maintain your wiki?

Language models do not get bored, do not forget to update cross-references, and can edit many files in one pass — the maintenance burden that makes humans abandon personal wikis is the part machines handle well.

Q12: Why is a knowledge base structured for humans expensive for an agent to query?

It forces the agent to read a great deal to answer anything. Atomic notes, typed edges and indexed summaries let it traverse instead of read — one demonstration drops a query from roughly nine thousand tokens to a few hundred.

Q13: What is the "harness" layer in the Infinite Brain methodology, and what problem does it solve?

A layer above individual brains that keeps them isolated for safety while promoting reviewed, validated ideas into shared canon and routing work by trust level — collaboration without shared-memory contamination.

Q14: What is the Open Knowledge Format, and what problem does it solve?

A vendor-neutral Google Cloud specification (v0.1, June 2026) representing knowledge as a directory of Markdown files with YAML frontmatter, one file per concept, cross-linked with ordinary links — so knowledge is portable between producers and consumers.