The Research Logbook

Acknowledgement Thanks to my colleague and friend Franco to go down the rabbit hole with me.

TL;DR. I spent an embarrassing amount of time reading how seventeen scientists — from Faraday to Tao — actually organised their work. None used zettelkasten. None used PARA. They all kept some variant of a dated logbook and a few supporting structures. But they also had something modern researchers don't: someone else handling the operational side. I distilled what they did for inquiry, added what's needed for delivery, and unified them into what I call the Research Logbook. Here it is (Oh, I plan to put it in an opensource app with AI hooks because why not).

How this started (badly)

That thing where you reorganise your notes instead of doing your actual research? That, but for months. Half-dozen half-finished Obsidian vaults. Tried zettelkasten, PARA, BASB, Logseq, Tana, briefly Mem even Capacities and Anytype. Same arc: aesthetic honeymoon, three weeks of joy, collapse under metadata weight. The usual AuDHD struggle between wanting structure but creating chaos. Eventually asked the obvious dodged question: what did people who actually produced great research do before any of this existed?

Why the popular systems don't fit

The systems I kept trying — zettelkasten, PARA, building-a-second-brain — are not bad ideas. But they solve a different problem than the one I have.

Luhmann's zettelkasten was built for synthesis from reading. If your work is connecting ideas across hundreds of texts, building arguments from literature, constructing theoretical frameworks from other people's writing — the atomic, interlinked note is a genuinely powerful unit of work. Social scientists, philosophers, historians, writers: zettelkasten was made for you.

But I am not building arguments from reading. I am running experiments. Experimental work has a fundamentally different structure. It is temporal: what I did, in what order, what I observed, what surprised me, what I tried next because of what I saw. An observation ripped from its sequence loses half its meaning. "Validation loss spiked at epoch 40" is useless without "because I had just doubled the learning rate after noticing a plateau at epoch 35." The context is the content. You need the log, not the card.

Mathematical and theoretical work fits even worse. A proof is not a collection of atomic ideas — it is an architecture. The pieces have internal dependencies, a logical order that carries the argument. Decomposing a derivation into interlinked notes does not make it more accessible; it makes it less coherent.

PARA solves the same underlying problem with different furniture — categorical folders instead of interlinked cards, but still oriented around filing things you consumed. And neither zettelkasten nor PARA handles the other half of modern research: delivery. PARA explicitly delegates task management to GTD. Zettelkasten says nothing about it at all. If you are a researcher in 2026, you are expected to bolt together a note system, a filing system, and a task manager, and hope they stay in sync. Three parallel systems for one job.

The question was never "which note-taking app?" It was: what does a system look like when you need to both investigate open questions and deliver on hard deadlines — without running three tools in parallel?

The rabbit hole

I looked at seventeen of them. Faraday, Darwin, Gauss, Euler, Hamming, Knuth, Grothendieck, Wolfram, Tao, LeCun, Bronstein, Turing, Einstein, Hawking, Schrödinger, Feynman, Erdős. Different centuries, different fields, wildly different personalities. I was looking for two things: how they organised their thinking, and how they managed their obligations. The first question had clear answers. The second was more revealing than I expected.

How they organised their thinking

What they did clusters into about five patterns:

Faraday kept a chronological laboratory diary from 1820 to 1862 — 16,041 sequentially numbered entries, with cross-references layered on top. He just wrote stuff down in the order it happened. That is it. That is the system.

Darwin kept "portfolios" — folders, basically — one per question he was working on. He dumped clippings, letters, his own notes, anything relevant, into the right portfolio. When he needed to write a chapter, he pulled the portfolio and worked through the accumulated material.

Hamming kept roughly ten to twenty important open problems in mind at all times, and dedicated his Friday afternoons to reflecting on what the important problems in his field were. He also famously (and provocatively) challenged colleagues at Bell Labs with the same question — pointedly enough that he got kicked off their lunch table.

Tao matches tasks to his fluctuating energy levels rather than forcing a rigid schedule, maintains a public blog as a thinking-out-loud system, and writes papers using a "rapid prototyping" method — a skeletal first draft covering the entire argument before filling in detail.

Wolfram has instrumented every aspect of his life — every keystroke, every step, a screenshot every minute, over 100 million keystrokes logged, a complete email archive back to 1989. The more routine the practical aspects of his life, the more energy freed for actual thinking.

Hawking, forced by ALS to abandon writing, developed visual methods — seeing equations as geometry, holding entire arguments as mental images of curves and surfaces. Werner Israel compared the achievements to Mozart composing an entire symphony in his head. Kip Thorne suggested the constraint may have made his geometric intuition uniquely powerful.

Grothendieck left 28,000 pages of handwritten and typed notes at the University of Montpellier — typed on poor paper, reinserting pages to use every corner. His filing system was non-existent. His mathematical architecture — schemes, toposes, motives — was so powerful that external organisation was irrelevant. The theory itself told you where everything belonged.

Feynman once created a notebook titled "NOTEBOOK OF THINGS I DON'T KNOW ABOUT" — a one-time exercise during his grad school qualifying exams, but widely cited as exemplifying his learning philosophy. His diagrams were a representational breakthrough that reorganised how an entire field structures its calculations.

And the rest — Gauss with his radically compressed Tagebuch of 146 terse entries, Euler producing 800+ papers seemingly from memory alone, Erdős carrying his mathematical life in two suitcases and distributing knowledge management across 500+ collaborators, Einstein working through problems via thought experiments and letters to Besso and Grossmann, Schrödinger deliberately jumping between fields until his cross-domain analogies produced wave mechanics in a six-month burst at 38, Knuth inventing literate programming so code and explanation became the same artefact, Turing leaving behind one surviving notebook of blunt candour ("I don't like this", "rather abortive", "ugly"), LeCun and Bronstein organising around research programmes and collaborator networks rather than any filing scheme.

The thing that surprised me most is how unsexy this is. There is no clever filing scheme. There is no addressing system. There is no app. The most consistent observation across all of these scientists is that the filing system is the least interesting part. The work is what mattered.

How they managed their obligations

Here is the second surprise: not a single one of them ran their own project management system. They either had someone else do it, eliminated the need, or operated in an institutional structure that handled it for them.

They delegated. Grothendieck had Dieudonné — who served as dedicated scribe and project manager for the EGA, rewriting Grothendieck's rough drafts into publishable form while Grothendieck focused purely on the mathematics. Hawking relied on a succession of graduate assistants and a personal assistant (Judith Croasdell, for roughly twenty years) who managed his schedule, travel, and public engagements. Darwin had Hooker as his operational collaborator — over 1,200 letters across forty years — and Lyell as intermediary with his publisher. Euler dictated papers to his sons and students after losing his sight.

They eliminated. Feynman called himself "actively irresponsible" — he refused every committee, every administrative role, every obligation that was not his own research or teaching, so that someone else would handle it. Knuth quit email in 1990 and replaced it with batch-processed physical mail, read roughly once every three months. Einstein at the Institute for Advanced Study had zero obligations: no teaching, no deadlines, no deliverables.

They centralised through sheer force. Wolfram functions as a deliberate one-person bottleneck at Wolfram Research — reviewing and directing all work personally, using live-streamed working sessions as his coordination mechanism. Gauss imposed a rigid seasonal structure: summers for geodetic fieldwork, winters for computation, and simply did not publish until results met his standard (pauca sed matura).

They outsourced to the network. Erdős distributed task management across 500+ collaborators — he arrived with notebooks of open problems, worked until a result emerged or he moved on, and left the coordination to whoever he had visited. Tao maintains a mental queue of problems at various stages, coordinating with collaborators via email and shared LaTeX drafts. LeCun and Bronstein run standard academic group models with regular meetings and high researcher autonomy.

They had structure imposed on them. Euler was salaried by the academies of Berlin and St. Petersburg, which expected regular memoirs — a soft but steady external deadline. Schrödinger had a standard teaching load; his burst of wave mechanics happened during a Christmas holiday. Turing nominally led Hut 8, but Hugh Alexander handled the day-to-day operations because Turing had little interest in running the section.

The pattern is unmistakable. These scientists had brilliant methods for inquiry — and the operational burden was carried by someone else. Dieudonné tracked EGA. Hooker coordinated Darwin's network. Hawking's assistants managed his calendar. Feynman just said no.

The modern researcher — especially early-career — does not have a Dieudonné. You are your own project manager, your own secretary, your own grant administrator. You file your own ethics applications, track your own conference deadlines, coordinate your own collaborators, and still need to do the actual research. The inquiry methods from the rabbit hole are brilliant, but they are incomplete. You need the operations track too, because nobody else is running it for you.

The method

Modern research is two jobs at once: inquiry (open-ended investigation driven by questions and evidence) and operations (delivery against deadlines, coordination with collaborators, accountability to yourself and others). No existing system handles both. The Research Logbook does, using two tracks that share a common substrate — the daily log, the weekly review, and a single set of projects that bridge the two.

Track 1: Inquiry — how you investigate

1. The chronological log (from Faraday)

What. A single, append-only stream where you record what you did, what you observed, and what you thought — in the order it happened. One file per day, date-stamped entries. You never go back and edit past entries; if you change your mind, you add a new entry that references the old one.

Why. It solves three problems at once. First, it eliminates the "where do I put this?" decision — everything goes in the same place. Second, it preserves context: when you look back at an experiment from three months ago, you do not just see the result, you see what you were thinking at the time, what you tried before, what you expected. Third, it creates an honest record. Researchers are prone to hindsight bias. The log keeps you honest.

The log is unified across all contexts. "Tuned the regularisation weight on the surrogate model" and "called the plumber about the leak" both go in the same day. Your day is unified; your log should be too.

How. Use whatever medium has the lowest friction. Each entry should contain, at minimum: what you set out to do, what you actually did, what you observed (including surprises), what you think it means, and what you plan to do next. For computational work, log the reasoning, not just the metrics: "I increased the learning rate because the loss curve looked flat after epoch 50 and I suspected the optimiser was stuck."

At the end of each week, add a few keywords per entry so you can search by topic later. That is Faraday's subject index, simplified.

2. Evidence portfolios (from Darwin)

What. A folder for each active research question into which you accumulate everything relevant: paper excerpts, experimental results, theoretical notes, counterarguments, half-formed ideas. Whatever applies.

Why. The log interleaves observations about many different questions. When it is time to write a paper or make a decision, you need all the evidence on a single question in one place. The portfolio solves this aggregation problem.

How. Create one folder per active question. Name the folder descriptively — sparse-vs-dense-ood is fine — but phrase the question it answers at the top of the folder's index note. "Sparse models" tells you the topic; "Does the sparse variant outperform the dense baseline on out-of-distribution data?" tells you when you are done. Think in questions, not topics — but keep the filesystem practical.

Inside each portfolio, do not worry about internal organisation. Just accumulate material. The key discipline is filing things as they arise — when you finish reading a paper, write a short note and drop it in. If you wait, you will forget your reaction, which is the most valuable part.

3. Important questions (from Hamming)

What. A short list of the questions that matter most to you right now. Not tasks. Not to-dos. Questions: things that, if answered, would genuinely change your situation.

Why. Most people drift. They start with an important question, get pulled into a side project, spend three months optimising something that does not matter, and look up to find that a year has passed. The important questions list is an attention-directing tool. It does not organise your notes — it organises your focus. It also sits above projects: a question like "Can we reduce sample complexity by an order of magnitude using learned simulators?" might spawn multiple projects over time. The question persists even as projects finish or get shelved.

How. Maintain two parallel lists. Research problems (3-5 items): "Can we reduce sample complexity by an order of magnitude using learned simulators?" Life priorities (3-5 items): "What does a sustainable exercise routine look like for my schedule?" Phrased as questions, not topics. Review both monthly. Are these still the right questions?

Track 2: Operations — how you deliver

Research questions and projects are related but not the same thing. A research question is open-ended — "Does X outperform Y?" It might take months, might branch, might turn out to be the wrong question. A project is bounded — "Submit the sparse-vs-dense paper to NeurIPS by May 17." It has a deadline, deliverables, and people counting on you. One research question can span multiple projects. One project can draw on multiple portfolios. A question can exist without a project (you are just curious), and a project can exist without a research question ("organise the lab retreat"). Keeping them separate prevents deadline pressure from polluting open-ended inquiry, and prevents open-ended inquiry from dissolving operational commitments.

4. Project maps (from Knuth, informed by all of them)

What. A lightweight, mutable one-pager for each active piece of finite work — the single page you open when you sit down after a two-week gap.

Why. This is where tasks live. Not in a global todo list, not in a standalone task manager — embedded in the project that gives them meaning. A project map answers five questions instantly:

• Where was I? (3-5 sentences of current state.)
• What do I believe and what would change my mind? (Current hypothesis or strategy. For non-research projects: what is the plan and what are the risks?)
• What are my next actions? (1-3 concrete tasks, tagged by energy level. This is the task surface for daily orientation.)
• Is anything time-sensitive? (Links to the commitments register.)
• Where is everything? (Links to portfolios, log entries, external resources.)

Tasks are born in the log ("I should try X next"), and if they matter beyond today, they get promoted to the project map's next-actions list. The weekly review checks: done? still relevant? replace with something better? The cap — maximum three next actions per project, maximum five active projects — means you are looking at fifteen tasks at most. That is a ceiling, not a backlog.

What a project map should not contain: a complete history (that is what the log is for), a long checkbox list (every time you open eighteen unchecked boxes, your brain says "I am failing"), or detailed timelines (research does not work that way).

How. Maximum five active projects across all of life. Everything else is parked or waiting. The cap forces prioritisation.

5. Stewardships

What. Simpler dashboards for perpetual responsibilities — health, finances, relationships, household — that never "finish."

Why. The critical distinction: projects end, stewardships do not. If health and finances occupy project slots, you permanently lose two of your five active slots to things that never finish. That leaves three for work that actually advances toward a goal. Separating them prevents perpetual responsibilities from silently crowding out finite work.

How. Stewardships run in the background with periodic commitments (annual tax filing, quarterly check-up, monthly budget review) that feed into the commitments register. They get a quick scan during the weekly review but do not compete for project slots.

6. Commitments register

What. A single, flat, date-sorted list of everything with a hard external deadline, regardless of which project or stewardship it belongs to. Conference submissions, visa applications, tax filings, a collaborator's thesis defence, a lease renewal.

Why. This is not a todo list. The distinction matters more than I expected. A todo is something you decided to do. A commitment is something someone else is counting on. Mixing them means the commitments drown in the todos and you break promises you meant to keep. Hard deadlines come from many different projects and stewardships; scattering them means you have to check every project to know if something is due. The register puts them all in one view.

How. One markdown file, sorted by date. Each entry has: the deadline, what it is, which project or stewardship it belongs to, and any preparation milestones with lead times. "Submit grant application — deadline October 15 — need draft by September 15 — co-PI feedback by October 1 — final revision by October 10." Those milestones feed back into the relevant project maps as next actions at the right time. The weekly review includes a forward scan: what is due in the next two to four weeks, and am I on track?

7. Energy-matched scheduling (from Tao and Wolfram)

What. Matching your tasks to your current cognitive state rather than forcing a fixed schedule.

Why. Research involves tasks at wildly different cognitive costs. Deriving a proof, debugging a subtle numerical issue, writing a paper introduction, answering emails — these require completely different types of mental energy. Forcing deep theoretical work when your brain wants something mechanical is not discipline; it is waste. The alternative to deep work on a low-energy day is not discipline — it is paralysis.

How. Tag tasks by cognitive level: deep (theoretical derivations, novel design, difficult writing), medium (running experiments, paper drafting, budget planning), light (email, admin, config work, routine errands). When you sit down to work, check your energy and choose a task that matches. Do not feel guilty about doing light work on a low-energy day.

The one exception: if you never have deep-work days, the problem is environmental. Protect blocks of uninterrupted time. Einstein sailed. Knuth quit email. You can block mornings, close Slack, and say you are unavailable.

The bridge: how the two tracks talk to each other

The log feeds both tracks. An experimental observation goes into the portfolio (inquiry). A realisation that "we need to rerun this before the deadline" becomes a next action in the project map (operations). The weekly review governs both: retrospective (what did I learn?) and forward scan (what is bearing down on me?). Projects reference the portfolios they draw on. Commitments decompose into milestones that surface in project maps.

The daily orientation is where the two tracks meet: you check your commitments (what is urgent?), glance at your project maps (what is important?), check your energy (what is realistic?), and pick one thing to start with.

The rhythm that keeps it alive

None of this works without review. Three touchpoints, deliberately minimal (this is where AI will makes our lives easier later on):

• Daily orientation: check the commitments register, glance at your project maps, check your energy, pick one thing to start with.
• Weekly review: read this week's log entries. Write a wins list (celebration first — this matters for ADHD brains). Update project map states. Quick-scan stewardship dashboards. Forward scan: what is due in the next two to four weeks? Choose one focus for next week.
• Monthly review: review whether your important questions are still the right ones, check whether portfolios are growing or stale, adjust project slot allocation.

Each review level reads material that already exists. None requires generating new analysis from scratch. The weekly review starts with wins because, for an ADHD brain, shame is the enemy of systems. If every review starts with "here is what you failed to do," you will stop reviewing. If it starts with "look what you actually accomplished," you will keep coming back.

The weekly review is also your Hamming Friday — the moment to step back and ask: am I still pointed at the right questions? Is this project still the best use of a slot?

Why this works for an ADHD brain

Two properties run through the entire design:

Decision-free at capture time. Open today's log entry, write it down, done. No "where does this go". No tagging. No linking. The cognitive cost of capture is as close to zero as a structured system can get.

Survives neglect. Disappear for two weeks — paper deadline, wedding planning — and you come back to a chronological log with a gap. Nothing to "catch up on", no inbox of 400 unprocessed notes glaring at you. You just start a new entry. Your project maps might be stale, but they are one-pagers — two minutes of reading and you know where you were.

Every system I tried before punished absence. This one does not.

Try it now (you do not need a tool)

The whole system is plain text. Here is what it looks like on disk:

research-logbook/
  log/
    2026-04-11.md
    2026-04-12.md
  portfolios/
    sparse-vs-dense-ood/
      index.md
      lecun-2024-sparsity-results.md
      experiment-log-april.md
    commute-vs-budget/
      index.md
  projects/
    neurips-submission/
      neurips-submission.md          # project map
      tasks/
        draft-section-3.md
        rerun-ablations-new-seed.md
    lab-retreat/
      lab-retreat.md
      tasks/
        book-venue.md
  stewardships/
    health.md
    finances.md
  important-questions.md
  commitments.md

You do not need all of this on day one. Start with three things:

1. Make a log/ folder. Create today's dated .md. Write what you did.
2. Make an important-questions.md. Three to five questions that matter. Re-read on Mondays.
3. Make a commitments.md. Promises to others, with dates. Look at it before saying yes to anything.

That is 80% of the value. Portfolios, project maps, and stewardships can wait until the habit sticks.

What is next

In Part 2 I will talk about Cuaderno — the tool I am designing to support this method. A CLI written in Rust, an MCP server for agentic AI flows and (maybe) a desktop app interfacing the CLI to a Tauri UI. Plain markdown on disk, that's non negotiable.

But try the method first. If it does not work with a text editor and three folders, no tool will save it.

-- Agustín