The Second Brain Builders

How to read this. Every claim is sourced to something public; where documentation is thin, the profile says so rather than filling the gap. The headline before the profiles: six of fifteen have a documented agent-facing system — Karpathy (a published pattern, not a build), Cherny (scoped to code), Wes Roth (a full build of Karpathy’s pattern — his profile was originally published as a false negative and carries a correction explaining the error), Reitz, Brier, Ghelbur. The rest are human-facing, several explicitly pre-LLM. That ratio is itself the finding: this design space has few predecessors, and the ones that exist agree more than they disagree.

---

Tier A — Big tech & AI labs

Andrej Karpathy

Founding member of OpenAI, ex-Tesla AI director, now Eureka Labs. He documents two systems he never connects in writing.

The daily driver is one plain-text Apple Note titled “notes” — no folders, no metadata beyond prefixes like “watch:” / “read:”. Items append to the top; retrieval is Ctrl+F; periodically he rescues still-relevant items back to the top and lets the rest sink. No AI.

The LLM Wiki gist (~Apr 2026) is the other, and he is explicit that it “describes the idea, not a specific implementation.” Three layers: raw sources (“immutable — the LLM reads from them but never modifies them”), the wiki of generated entity and concept pages (“The LLM owns this layer entirely”), and the schema — a CLAUDE.md/AGENTS.md that “you and the LLM co-evolve.” Three operations: ingest, one source at a time (“a single source might touch 10-15 wiki pages”); query; and lint, a watchlist of contradictions, stale claims superseded by newer sources, orphan pages, missing cross-references. Navigation is index.md (every page, one-line summary) plus an append-only, grep-able log.md.

Lessons. “Obsidian is the IDE; the LLM is the programmer; the wiki is the codebase.” Folders cost “way too much cognitive bloat.” “Good answers can be filed back into the wiki as new pages.” “The human’s job is to curate sources… The LLM’s job is everything else.” On scale: index.md works at “~100 sources, ~hundreds of pages” and “avoids the need for embedding-based RAG infrastructure.”

Honesty. The lint categories are named, not codified — no algorithm, thresholds, or script. No example schema file, no directory tree, and no page count from his own real wiki exists publicly.

Boris Cherny

Creator and Head of Claude Code at Anthropic. His documented system is Claude Code used on itself: a git-checked-in root CLAUDE.md the team edits multiple times a week (“Anytime we see Claude do something incorrectly we add it”), ~2.5k tokens as of Jan 2026; a separate auto-memory layer the model writes itself; and auto-dream, a subagent that periodically consolidates it, merging duplicates and dropping stale entries, likened to REM sleep. He went from ~5 parallel sessions in Jan 2026 to “thousands” of agents on busy days by June.

Lessons. “Ruthlessly edit your CLAUDE.md over time.” Then a partial reversal in Aug 2026: “Every 6 months, delete your CLAUDE.md, delete your skills, delete your hooks” — and, on prepping Opus 5, “We deleted 80% of the system prompt,” after which the model got better.

Honesty. An engineering-repo memory system, not a life second brain — no source shows personal notes, reading, or journaling, and none uses the phrase “second brain.” The Jan and Aug quotes are well corroborated; the auto-memory/auto-dream wording comes only from a self-declared fan site that paraphrases, the circulated “Boris Cherny’s CLAUDE.md” gist is an unattributed third-party mirror, and one teardown found auto-dream behind a disabled flag — so whether he runs it personally is unconfirmed.

Simon Willison

Creator of Datasette, co-creator of Django. His system is public writing that doubles as memory: a main blog since 2002 for polished work, plus til.simonwillison.net — a deliberately low-barrier TIL site, markdown in a public repo, 150+ topic folders, 581 entries, each typically assembled in 10–15 minutes. Underneath both, GitHub Issues as his primary work-notes system across ~250 projects, with threads over a hundred comments long that are entirely him; in May 2025 he counted 48,500 issues and comments. He deliberately separates “current” documentation from “temporal” documentation of what he was thinking on a date, kept in dated comments. Retrieval is public search — Google and GitHub — not private tags.

Lessons. “The bar for a TIL is literally ‘did I just learn something?’” “I’ve outsourced my memory to these external systems.” GitHub Issues is “almost the best notebook in the world” — with one named gap: no synced offline support, so he falls back to Apple Notes.

Honesty. Assembled from scattered posts 2021–2026; no single “here is my system” essay, and he never says “second brain.” One figure conflict, both his: 48,500 in print, “~45,000” on a podcast weeks later. AI builds the plumbing, not the memory — Claude wrote his stats query and built his Feb 2026 “beats” feature, but no source describes an agent that files, tags, or retrieves his notes.

Stephen Wolfram

Founder of Wolfram Research; documenting his personal infrastructure since 2012. A filesystem split into an “active” part synced across every machine and an archival part on a fileserver, with few top-level folders under a fine-grained taxonomy (named examples: Events, subfoldered by year then event; Designs). Retrieval is a single search box on his homepage — a “metasearcher” hitting ~30 years of email, every file, and every scanned document at once. Scale (2019): ~815,000 emails sent, ~2.3M received since 1989; 1.7M files backed up to 1980; ~230,000 scanned pages. All built and queried in his own Wolfram Language, alongside keystroke logging and near-hourly screen captures.

Lessons. On total logging, after one catastrophic crash: “it’s cheap to just record everything like that.” On clutter: “any flat surface represents a potential ‘stagnation point.’”

Honesty. No LLM layer is documented — a checked absence, not a gap: his 2023–2026 AI writing is product work for Wolfram Language users, never described as layered onto his own search box. The primary essay could not be fetched (TLS error; the Fast Company excerpt 403’d), so this is reconstructed from consistent third-party summaries and a podcast transcript — scale figures corroborate, but quoted wording is second-hand.

Steven Johnson

Author of 14 books, Editorial Director of Google Labs, co-creator of NotebookLM — the cleanest case of an old corpus given a new query layer. Twenty years of highlights, reported at 7,000 quotes (Feb 2024) growing to 8,000 (Sept 2024): about 1.3 million words across fifteen documents. Pipeline: Kindle → Readwise → Export to Docs → Google Docs → one NotebookLM notebook named “Commonplace Book.” Consolidating into fifteen giant files rather than one per book is a deliberate workaround for the per-source word cap; per-quote author/title/page metadata is what lets answers cite back into context. His documented test — “give me two quotes: one about cities from Jane Jacobs and another about ant colonies from Deborah Gordon. explain a surprising connection between them” — returned both correctly cited, with a connection “not far at all from the argument” he published years earlier in Emergence: 11 seconds versus 5 minutes 8 seconds by hand.

Lessons. The tool exists “to support genuine thinking rather than create illusions of understanding.” Consolidate aggressively to fit source limits; attach metadata at ingest so the system can cite rather than paraphrase.

Honesty. The wording of NotebookLM’s answer is unrecoverable — only his paraphrase. His X threads and a paywalled interview were inaccessible, and no taxonomy inside the fifteen documents is confirmed — only that they are large flat files with per-quote metadata.

---

Tier B — Big creators

Wes Roth

AI-news YouTuber (~320k subscribers), operator of NATURAL 20. Architecture: documented — this profile originally said otherwise and was wrong (see the correction below). His system is the most complete public build of Karpathy’s LLM Wiki pattern in this report: an Obsidian vault run by Claude Code, documented in his video “Claude Built the Ultimate Second Brain” (2026-07-13; the vault tour runs ~16:23–19:06) and a written companion guide on natural20.com.

The build: a CLAUDE.md rulebook Claude Code reads before acting (“the rulebook Claude Code reads before it works”); Raw/ for immutable original sources; Inbox/ for quick captures plus an automated daily news digest that is not auto-ingested — he curates what enters; and Wiki/ with Index.md, Log.md, and Entities/Concepts/Summaries folders. Conventions: wikilinks throughout, YAML frontmatter on every note (type, dates, tags), absolute dates only, S - /R - filename prefixes, and wiki claims must cite their summary page. Claude Code runs Karpathy’s three operations — ingest (one source ripples through ~5–15 entity/concept pages), query (Index first, answers cite pages, wiki knowledge distinguished from model knowledge), lint (contradictions, orphans, stale claims) — plus a boundary rule: never modify Raw, deprecate rather than delete.

Lessons. “The brain in your skull is brilliant at thinking and terrible at storage.” He repeats the Karpathy framing: “Obsidian is the IDE, the LLM is the programmer, and the wiki is the codebase.” (Quotes are from his written guide; a verbatim video transcript was not retrievable.)

Correction (2026-08-15) — and what it teaches. The first published version of this profile called Wes Roth “a confirmed negative result: no documented system.” That was wrong, and the error is more instructive than the profile. The video and guide existed a month before this research ran. The original profiler missed them for two compounding reasons. First, modality bias: its searches were text-shaped, and his documentation is primarily a video — it never enumerated his channel’s actual upload titles, one of which is literally “Claude Built the Ultimate Second Brain.” Second, and worse, skepticism anchoring: the profiler correctly caught one real misattribution (an “Obsidian + Claude Code” walkthrough credited to Roth that is actually by AI Andy, Jun 16 2026) — and then classified every adjacent signal as more of the same fabrication. It had Geeky Gadgets’ coverage of Roth’s own video in its source list and dismissed it as “media credit only” because the byline was a staff writer’s. One true debunk became a license to stop looking. The no-embellishment rule worked exactly as designed on the fabricated “knowledge graphs” claim; the leap from “some claims about him are fabricated” to “no system exists” was the profiler’s own embellishment — in the negative direction. Corrected after Aaron supplied the primary source.

Tiago Forte

Author of Building a Second Brain, creator of PARA, now pivoting to “Personal Context Management.” PARA is organized by actionability, not topic, and is still his framework in 2026. Evernote for roughly a decade and still an active personal connector in June 2026; a dedicated Inbox notebook triaged into PARA; Readwise (30,000+ highlights since 2018); and RandomNote, a tool he built to surface random notes from Projects/Areas/Resources but deliberately excluding Archives, used 20–30 times a day. The 2026 AI layer is the concrete part: a published audit of his own Claude connectors, a “Master Prompt” standing context document, and the principle that every category of data gets one canonical home. His April 2026 evidence: 222 files left to pile on his desktop for a month, then hand-sorted into PARA in 36 minutes — the point being not sorting speed but that an agent needs a coherent folder to work from, after a failed attempt to have Claude draft a proposal from material scattered across Drive and Notion.

Lessons. “Put the note in the first place it’s going to be used.” “I want to remember as little as possible, not as much as possible.” “You’d want each kind of data in your life to have a canonical ‘home.’” “Personal Context Management is replacing Personal Knowledge Management.”

Honesty. He test-drove Obsidian, Tana, and Mem in Nov 2023 and rejected all three; a separate import of “thousands of notes” into Notion was an experiment, not a migration. Things 3 and “Focus Days” appear only in a 2019 interview, and Progressive Summarization is unconfirmed as current practice. He states plainly that multi-connector AI dashboards still required substantial human interpretation to be useful.

Nat Eliason

Ran the paid Roam course during the 2020–21 PKM boom; now building Founders School. The 2020 system (“V4”): Roam as central database, chosen because standard apps “work kind of like a filing cabinet, in that each note goes into one and exactly one place.” Daily Notes as the inbox — “Most of my capture in Roam goes straight into the Daily Note” — with linking done progressively rather than at capture: “notes live nowhere and everywhere… no note lives inside another.” Plus a PARA adaptation, a person-CRM auto-aggregated via linked references, and 250+ book notes. No AI; this predates it.

The arc. In Aug 2022 he wrote “Don’t Put the Tool Before the Craft”: “I’ve abandoned most of the productivity techniques I blogged and talked about over the last decade.” The replacement was three lines — tasks in Things, notes in Apple Notes, “try to get a few satisfying things done per day.” His diagnosis: “Productivity obsession is often a symptom of deeper dissatisfaction,” and tool-switching is the trap — “I could spend weeks practicing knife skills and it would only save me a few minutes every week.” From the Roam era, defending mess: “cleanliness and the beautiful formatting and layout actually gets in the way of making things intuitive.”

Honesty. Both halves are in his own named, dated writing — unusually strong sourcing for a “he abandoned it” claim. The phrase “note-taking is dead” is sometimes attributed to him; no instance was found. His 2023 a16z PKM podcast has no retrievable transcript, so nothing is quoted from it.

Marie Poulin

Notion consultant and educator; runs “Notion Mastery.” Two top-level workspaces — Marie HQ (personal) and a shared business one — PARA-derived but explicitly loose: “I take a lot of liberties with PARA and don’t use it literally.” Two moves matter for Notion v2. First, areas are pages, not databases: “Each of these areas on the side here are pages now and not databases,” with the underlying database embedded beneath in several filtered views, and an icon convention encoding type at a glance (solid circle = database, open circle = page). Second, deliberate duplication: “Lots of databases actually appear in multiple places in my workspace. It’s a way for me to not forget what is most important.” And dashboards, not PARA categories, are the real navigation layer — “These days I very rarely go and click on an individual database” — by Aug 2025 crystallized into a single “Today” page where she spends “90% of my time.”

Lessons. “There are no rules.” “The most effective systems are co-created, not adopted wholesale.” On change: “This isn’t a failure – it’s a feature.” On capture: “What is this in service of? How is my future self benefiting from this information?”

Honesty. No AI adoption, and her latest stated position is aversion — her Jan 2026 review describes “extreme AI fatigue” and says “AI is making me want to spend more time in the garden.” The most recent full workspace tour is still the 2020 post, and the often-cited “~15 dashboards” could not be verified — her public template lists 9.

Nick Milo

Founder of Linking Your Thinking; teaches Obsidian PKM and sells pre-built vaults. Top level is ACE: Atlas (ideas and knowledge, by relatedness — where Maps of Content live), Calendar (dailies, meetings, journals), Efforts (projects by intensity). Maps of Content are the mechanism and the explicit alternative to folders: hub notes linking a cluster, with commentary around each link. A note can live in multiple MOCs or none, while “a note can’t be in two folders at once, nor can you add headers or other structure to folders” and “folders carry no context beyond their name.” His trigger heuristic is the “Mental Squeeze Point” — when unsorted knowledge becomes “so messy it overwhelms and discourages you,” that is when you build the MOC. ACE replaced a six-part scheme (ACCESS), which he had framed as provisional: “not the gospel. It’s meant for you to tear apart and make your own.”

Lessons. On leaving Evernote: exporting broke his links and formatting, and “even years later, Evernote still manages to punch me in the gut.” On collecting: clipping articles was “increasing the NOISE and decreasing the SIGNAL.” On structure: “structure must be earned” — pre-structuring a vault in one sitting “leads to a fragile building that will break in a week.”

Honesty. He publishes no naming or tagging convention — the ACE post explicitly declines to give one. His AI product (“Linking Your AI,” an “AIOS” with a me.md file) names eight sub-systems but gives no worked example, prompt text, or folder diagram for any — packaging, not documented mechanics. The ~1,000-note scale figure describes the vault he ships to customers, not his own.

---

Tier C — Practitioners

Kenneth Reitz

Creator of Python Requests; documents a large Obsidian vault worked on by Claude Code across an essay series. March 2026: 467 notes, Obsidian Git, numbered Dewey-style folders (000 Meta … 900 Archive), every note carrying the same frontmatter (type, role, status, date, themes). At the root, a ~200-line CLAUDE.md he calls “an API contract,” with sections Vault Overview, Conventions, and What NOT To Do — don’t create files unless asked, don’t restructure folders, no emojis, don’t editorialize. He frames the use as “write-then-think,” asking questions no human runs manually: “What themes appear in March but disappear by June?”

The arc. By April: 786 files and the numbered folders were gone, replaced by meaning-named ones (Knowledge/, Writing/, Life/, Projects/, Daily/). He calls the numbering “training wheels”: “The structure didn’t change because I learned a better productivity system. It changed because my relationship to the vault deepened.” By June: ~4,800 files, “the three-million-word vault,” with maintenance logs like 1,446 em-dashes fixed across 81 files and ~890 broken wikilinks resolved.

Lessons. “The AI holds the architecture, the continuity, the syntax, the stamina. I hold the spark, the facts, the taste, the kill switch, and the life.” On upkeep: “the fiddling was doing a different, legitimate job.”

Honesty. All self-reported on his own blog, no external corroboration. The full 26-plugin list and verbatim CLAUDE.md were not retrievable, and quotes came through an automated fetch step — read them as close-to-verbatim. The 467 → 4,800 growth in three months cannot be independently confirmed.

Noah Brier

Co-founder of Percolate (acquired 2019), now Alephic. A local Obsidian vault of ~1,500 markdown notes synced to a private GitHub repo. The load-bearing decision: Claude Code is launched from the vault root, not a project subfolder — “I’m starting it in the full Obsidian vault” — so it can search the entire knowledge base. Project folders hold a Chats subfolder (clipped transcripts of his sessions with other models), a Daily Progress file of AI-written research summaries, a Research folder, and a Conclusions note workshopped iteratively. A “thinking partner” subagent carries an explicit ban on drafting: “your role is to facilitate thinking” and “don’t help me write anything right now.”

Lessons. On agents overreaching: “Even when you very specifically tell it not to try to do your work, it still often tries to do your work… you have to really be like, no, I said no.” On what’s underrated: “There’s entirely too much focus on its ability to write and not enough focus on its ability to read.”

Honesty. Strong sourcing — a full transcript plus a public MIT repo (heyitsnoah/claudesidian) with a numbered PARA tree and named skills. But that repo is explicitly a starter kit built by his firm, “inspired by” the interview — not his private vault, so those folder and script names are confirmed for the template only. The interview gives one illustrative project, not a map of the 1,500 notes.

Jerry Michalski

Technology analyst; has built “Jerry’s Brain” in TheBrain continuously since December 1997, with no migration ever documented. No folders, no fixed hierarchy: each “thought” is a short word or phrase, connections typed parent/child/jump, one active thought center-screen. Scale: 281,000 thoughts / 500,000+ links (2015); 572,000 / 1,120,000 (Mar 2024) — entirely manual at 50–60 thoughts a day. Two conventions do the work: a deliberate no-orphan rule (everything “hooked like a Christmas ornament onto some branch”) and minimal color coding. Retrieval is type-ahead search, not browsing — without it, TheBrain “would be half as valuable.”

Lessons. “An LLM can retrieve. It can’t curate with intention.” Adding to the Brain forces deliberate “system two” thinking — is this worth remembering, and what does it connect to. On the artifact: “I make no claim of completeness or accuracy.”

Honesty. No AI, by design — a hand-curated pre-LLM system, and he frames LLMs as a contrast to his practice, not a component. He doesn’t delete dead links; he tolerates rot and relies on the Wayback Machine. Two gaps: no source has him describing what he actually can’t find at 570K nodes, and “he never restructured” is an inference from silence across ~10 sources, not something he said.

Dmitrii Gerasimov (karlicoss)

Engineer; author of HPI (Human Programming Interface), Orger, and Promnesia — built for himself, not as products. Three layers. Export: per-service scripts that save data “as intact as possible,” no cleanup, ISO8601-timestamped. Local cache: raw JSON/SQLite/CSV, one snapshot per export, kept after services die (his Jawbone and Endomondo data outlived those companies). HPI: a Python namespace package (my.reddit.all) that normalizes raw files into typed objects at runtime, not at write time — “the modules contain no data, only code to operate on the data.” Consumption is hpi query <module>, streaming JSON with date filters, pipeable into jq. 40–50+ modules, with his caveat: “even though there’s a lot of stuff on the diagram, it’s still incomplete.”

Lessons. “It feels very wrong that things can be forgotten and lost forever.” “I want to have a choice whether to forget or remember events.” On databases: they “aren’t flexible enough to fit your data.”

Honesty. No AI, agents, or MCP anywhere in the sources — searched for specifically, not assumed. Pre-LLM by construction (2019–2020); its typed, JSON-streamable shape is exactly what would make it agent-queryable, but that is an observation about its shape, not a claim he makes. The public writing stops around Feb 2021 while the repo shows commits through August 2026 — the narrative froze; the system didn’t.

Eugeniu Ghelbur

AI Automation Engineer at Single Grain; author of obsidian-second-brain, an open-source Claude Code skill connecting an Obsidian vault to seven CLI agents. The closest analogue to what Aaron is building. 46 commands compiled from one source tree into seven agent builds, each checked in CI. Vault layout: _CLAUDE.md, index.md, log.md, SOUL.md, a CRITICAL_FACTS.md of ~120 tokens always loaded; raw/ for immutable sources; wiki/{entities,concepts,projects,daily,logs,reviews,tasks,decisions}. Ingest rewrites rather than appends — resolving contradictions across 5–15 pages per source, facts tracked bi-temporally (when true vs. when the vault learned it). Four scheduled agents: morning brief 8am, nightly consolidation 10pm (close day, reconcile, synthesize, heal orphans), weekly review Friday, health check Sunday — plus an opt-in background agent run headless with a narrow tool surface (Read/Write/Edit/Glob/Grep only, no shell, no network, add-and-update-only, never delete).

His maintenance spec, OKM (Open Knowledge Metabolism): every stored fact must be timeless, dated, or a pointer — slow-changing knowledge is stored; fast-changing facts link out with an “as of” stamp rather than being copied in to go stale. A linter enforces it. On his ~2,350-note vault, keyword recall@10 was 1.0 for exact terms but 77% for paraphrased questions, and non-English recall@5 went from 13% to 63% with a multilingual embedding model. His boundary: “below roughly 50,000 stored items, lexical search with a hard budget beats a vector database.”

Evolution. The changelog is forensic: v0.9.0 made the background agent inert by default after it had been arming itself on an env var a normal install already set; v0.11.0 added hybrid search only after an eval showed lexical-only topped out at ~17% recall on paraphrased queries; v0.12.0 added atomic note writes, because an interrupted write could previously zero out a note.

Lessons. On splitting knowledge from work: “the thing I spend the most hours inside, the actual codebase, lived nowhere near it,” so “the decision note and the code that the decision shaped had no idea the other one existed.” And the best documented payoff in this report: about to migrate a database from Railway to Supabase, Claude searched the vault and surfaced his own six-month-old rejection of Supabase over “cold start latency on serverless functions killing our API response times.”

Honesty. Repo stats and changelog dates were verified against the raw README and GitHub API. But this is product documentation written by its maintainer; his own ~1,700–2,350-note vault is the only one it is measured on.

---

What actually transfers

1. Capture undifferentiated; file later — but the drain is the unsolved part. Karpathy appends to the top of one note; Eliason’s capture “goes straight into the Daily Note”; Forte has an Inbox notebook; Milo has Calendar. Nobody successful decides where a thing goes at capture time — except Michalski, who does exactly that and considers the deliberation the point. For v2: the inbox is consensus. But everyone documents the inbox and almost nobody documents the drain — Ghelbur is the only builder here with an enforced, scheduled one. The scheduled drain is thus both the highest-leverage and least-validated piece of Aaron’s design: right idea, one precedent, so instrument it.

2. Structure gets simpler over time, never more complex. Milo went ACCESS (6) → ACE (3). Reitz abandoned numbered folders for meaning-named ones. Poulin converted areas from databases to pages. Cherny deleted 80% of a system prompt. Eliason collapsed a Roam graph into Apple Notes. Karpathy runs one file. Dissent: Wolfram maintains a fine-grained taxonomy at 1.7M files and has never simplified. For v2: flat-ish shallow is well supported, and the corollary is Milo’s — “structure must be earned.” Ship less than feels complete, let the steward propose additions, and assume v2 becomes v3.

3. Links carry the context folders can’t. Milo: “a note can’t be in two folders at once, nor can you add headers or other structure to folders.” Eliason: “notes live nowhere and everywhere.” Michalski’s no-orphan rule; Karpathy’s lint hunts “orphan pages with no inbound links.” For v2: wiki links are the right primary structure, and “no orphans” should be a steward-checkable rule — Ghelbur’s nightly agent literally “heals orphans.” Poulin adds a Notion-specific move worth stealing: deliberate duplication of the same view in several places “to not forget what is most important.” Redundant surfacing is not clutter.

4. A maintained index beats embeddings at this scale — with one measured caveat. Karpathy: index.md works at ~100 sources and “avoids the need for embedding-based RAG infrastructure.” Ghelbur numbers the boundary at ~50,000 items. Michalski, at 570,000 nodes, still relies on type-ahead search. For v2: keep an explicit index page the agent reads first and lean on Notion search — but know the failure Ghelbur measured: recall@10 was 1.0 for exact terms and 77% for paraphrased questions. If Aaron queries conversationally rather than with the words he filed under, that 23% is where things silently disappear.

5. Ingest must rewrite, not append. That is what makes it a wiki. Karpathy: “a single source might touch 10-15 wiki pages.” Ghelbur: 5–15 pages, contradictions resolved, facts bi-temporal. For v2: the filing agent’s job is not to create a page — it is to update the pages a new item touches. A filing agent that only appends produces a log, not a knowledge base. Ghelbur’s split between when a fact was true and when the vault learned it is the cheap version of this, and pairs directly with the freshness rule.

6. Every mature system has a garbage collector, scheduled and machine-run. Karpathy’s lint watchlist; Cherny’s auto-dream merging duplicates and dropping contradicted entries; Ghelbur’s four scheduled agents; Reitz’s counted maintenance logs. Karpathy’s reason for delegating: “LLMs don’t get bored, don’t forget to update a cross-reference.” For v2: the steward is well validated — give it Karpathy’s named categories as its checklist, the nearest thing to a shared standard here. Counter-note from Reitz: “the fiddling was doing a different, legitimate job.” Silent automation removes a form of contact with your own material; keep the steward’s output reviewable.

7. Deletion is the genuine disagreement — and it resolves on who. Karpathy never deletes; items sink “almost as if under gravity.” Michalski never deletes and tolerates link rot. Forte archives. Against them: Cherny, whose most emphatic recent advice is to delete the whole instruction layer every six months and see whether the model still needs it. And note where Ghelbur landed after his background agent armed itself accidentally: add-and-update-only, never delete. For v2: ambient and background agents should never delete; the scheduled steward proposes and a human confirms. Cherny’s advice applies to the instruction layer, where wholesale deletion is a test rather than a loss. Content and instructions deserve opposite policies.

8. The instruction file is the actual architecture. Reitz’s ~200-line CLAUDE.md is “an API contract” with a What NOT To Do section. Karpathy’s third layer is the co-evolved schema doc. Cherny edits his multiple times a week, then advises deleting it. Ghelbur keeps a ~120-token always-loaded core. For v2: the hosted MCP is the access layer, but the schema document is the design. Three borrowings: write the What-NOT-To-Do section first, because Brier’s failure mode is universal — “you have to really be like, no, I said no”; keep the always-loaded core small enough to be free; and run Cherny’s deletion test on a cadence, since a rule the model no longer needs is pure cost.

9. Freshness has one codified precedent — and everyone else’s failures corroborate it. Ghelbur’s OKM rule is nearly verbatim Aaron’s: timeless, dated, or a pointer, enforced by a linter. Corroboration that this is the right problem: Karpathy lints for “stale claims that newer sources have superseded”; auto-dream exists because memory gets “messy — outdated assumptions, overlapping notes”; Michalski’s link rot is the untreated case. For v2: well founded — and Willison supplies the missing distinction. He separates “current” from “temporal” documentation: notes about what he was thinking on a date, deliberately never updated. That differs from a dated fact that decays — it is correct forever because it is stamped. Aaron’s “dated” bucket should distinguish decays from is a record, or the steward will keep flagging journal entries as rot.

10. Curate inputs, delegate the rest — and give the agent the whole workspace. Karpathy: “The human’s job is to curate sources… The LLM’s job is everything else.” Reitz: “I hold the spark, the facts, the taste, the kill switch, and the life.” Brier’s thinking partner is banned from drafting outright. And Brier’s structural choice matters most for the MCP layer: he launches at the vault root, not per-project, so the agent can search everything — the argument for scoping the hosted MCP at workspace level rather than per-database. His related line belongs at the center of v2: “There’s entirely too much focus on its ability to write and not enough focus on its ability to read.”

11. Answers must be filed back. Karpathy: “good answers can be filed back into the wiki as new pages,” so explorations compound instead of vanishing into chat history. Brier does this with per-project Conclusions notes and AI-written Daily Progress files. Johnson notably does not — his answers stay in the chat and his corpus is read-only. v2 needs a defined home for agent-generated synthesis, distinct from captured raw material. Karpathy’s raw/wiki split is exactly that, and it is the one architectural line every agent-facing system here draws.

On English page titles. Aaron’s rule has almost no direct precedent: not one of these fifteen publishes a naming convention, and Milo’s ACE post explicitly declines to. The support is indirect — Michalski’s short plain-language labels, Karpathy’s grep-able log-line format, and Ghelbur’s measured multilingual collapse (13% recall@5 before a multilingual model). That last is the real argument: a title in a language the retrieval layer indexes poorly is a page that quietly stops existing. Treat English titles as a retrieval decision, not a style decision, and say so in the schema doc.

The disagreements worth keeping open. Michalski would say the filing agent is the mistake: “An LLM can retrieve. It can’t curate with intention.” Eliason built the most elaborate system in Tier B, abandoned it, and concluded that “productivity obsession is often a symptom of deeper dissatisfaction.” Poulin, a full-time Notion professional, reports “extreme AI fatigue” and has adopted none of this. Karpathy’s own daily driver is one text file with no AI in it. None of this invalidates v2, but it sets the bar: the system has to pay for its maintenance in retrievals that change a decision. Ghelbur’s Supabase moment is the only documented instance across these fifteen of a second brain preventing a concrete mistake — one worked example out of fifteen serious builders is a sobering base rate, and it is the outcome to instrument for.

---

Sources

Karpathy — gist.github.com/karpathy/442a6bf555914893e9891c11519de94f (LLM Wiki, ~Apr 4 2026); bearblog.dev post on the append-and-review note. Cherny — x.com/bcherny/status/2007179832300581177 (Jan 2 2026; 402, not fetchable) + twitter-thread.com mirror; howborisusesclaudecode.com (fan-made, paraphrases); infoq.com/news/2026/01/claude-code-creator-workflow/; gist.github.com/hqman/e29cb6386c539d795767e8c3fd2c959b; gist.github.com/joyrexus/e20ead11b3df4de46ab32b4a7269abe0; threadreaderapp.com/user/bcherny; antoniocortes.com (Mar 30 2026); tessl.io (auto-dream teardown); YC talk coverage (Aug 2026). Willison — til.simonwillison.net; simonwillison.net/2021/May/2/one-year-of-tils/; /2025/May/26/notes/; /tags/github-issues/; /tags/llm-memory/; Heavybit Generationship (Jul 2025); mtlynch.io interview notes (2025); “beats” post (Feb 2026). Wolfram — writings.stephenwolfram.com/2019/02/seeking-the-productive-life… (primary; TLS error, not fetched); “The Personal Analytics of My Life” (2012); “Analyzing Your Email with Mathematica” (2012); fastcompany.com/90432278 (403); news.ycombinator.com/item?id=26045380; Irreal; Jorge Arango; Tim Ferriss #637 transcript (Nov 2022). Johnson — “Writing At The Speed Of Thought,” Adjacent Possible (Dec 2023); “Introducing NotebookLM” (2023); “Capturing and Colliding”; Readwise NotebookLM export docs; Hard Fork (Sept 2024); his X threads (402); Exponential View interview (paywalled); “The Glass Box and the Commonplace Book” (2010, unfetchable). Wes Roth — youtube.com/watch?v=cwf2vEAigKA (“Claude Built the Ultimate Second Brain”, 2026-07-13 — the primary source, supplied by Aaron in the correction); natural20.com/using-claude-code-to-setup-a-second-brain-aka-llm-wiki (his written companion guide); youtube.com/@WesRoth; natural20.com + About; geeky-gadgets.com/build-ai-second-brain-2/ (coverage of his video — originally misread as mere media credit); dutchstartup.ai and youtube.com/watch?v=Y2rpFa43jTo (the genuinely misattributed AI Andy video, Jun 16 2026); course.natural20.com/authors/wes-roth; skool.com/@wes-roth-8854. Forte — fortelabs.com: PARA series and “P.A.R.A. Part 3” (Aug 2020), “Progressive Summarization,” Sept 2023 tool post, Nov 2023 Obsidian/Tana/Mem test, “Introducing the AI Second Brain” (Mar 2026), PARA-for-AI post (Apr 13 2026), “The Ultimate Guide to Claude Connectors” (Jun 29 2026); Superorganizers interview (Dec 2019); RandomNote Web (2019–2020). Eliason — nateliason.com/blog/roam (Jan 18 2020); Every/Almanack “How I Took Notes on 250+ Books in Roam” (Sep 1 2020); Every/Superorganizers “Nat Eliason is Living in the Matrix” (Apr 28 2020); blog.nateliason.com/p/tool-before-craft (Aug 1 2022); a16z “A True Second Brain” (Aug 31 2023, no transcript); current nateliason.com bio. Poulin — “Updated Notion Workspace Tour” (Aug 12 2020); “Zone 0: The Self” (Aug 10 2020); “The Notion dashboard I use every day” (Aug 26 2025); “The universal principles behind every productivity system” (Aug 14 2025); “2025 Year in Review” (Jan 30 2026); Ness Labs interview (Nov 26 2020); her public all-in-one template. Milo — linkingyourthinking.com and blog.linkingyourthinking.com (ACE Folder Framework, MOC posts, Mental Squeeze Point); Ideaverse Pro / Ideaverse Zero pages; Linking Your AI page; LYT Workshop curriculum; X, May 2022 (ACCESS “not the gospel”); Substack notes; YouTube @linkingyourthinking. Reitz — kennethreitz.org: “Obsidian Vaults and Claude Code” (2026-03-06), “Infrastructure for One” (2026-04-16), “Tending the Vault” and “How I Write Now” (2026-06-11), “A Framework of One’s Own,” “Breaking Changes.” Brier — every.to/podcast/transcript-how-to-use-claude-code-as-a-thinking-partner (2025-09-10); github.com/heyitsnoah/claudesidian; noahbrier.com/about; brxnd.ai/people/noah-brier; every.to/chain-of-thought/the-end-of-organizing (checked; does not mention Brier). Michalski — jerrysbrain.com/welcome; bra.in/jerry; TheBrain blog “Inside Jerry’s Brain” and the 18th-anniversary post (Dec 2015); TheBrain 2023 event page; LinkedIn “Behold, Stump Jerry’s Brains!”; The Informed Life ep. 112; Sketch Your Mind 2025 session page. Inaccessible: old.thebrain.com big-thinkers page (404), blogarchive.thebrain.com (DNS failure), forums.thebrain.com NPR thread (403), thrivingonoverload.com (403). Gerasimov (karlicoss) — beepb00p.xyz: “How to cope with having a fleshy human brain” (May 2019), “The sad state of personal data and infrastructure” (Dec 2019), “What data on myself I collect and why” (Jan 2020), “Against unnecessary databases” (Jan 2020), “Building data liberation infrastructure” (Feb 2020), HPI announcement (Mar 2020), Promnesia (May 2020), “Map of my personal data infrastructure” (Feb 2021); github.com/karlicoss — HPI (incl. DESIGN.org, commits through Aug 2026), Orger, Promnesia, cachew, myinfra. Ghelbur — github.com/eugeniu-ghelbur/obsidian-second-brain (raw README, CHANGELOG.md v0.5.0–v0.14.0, GitHub API stats verified 2026-08); The AI Operator (Substack): “How I gave Claude Code memory without a vector DB,” “HUGE update on obsidian-second-brain: The Architect,” “I built a tool that lets my Obsidian vault argue with me.”