offline
Marcus Obst
webworker.js

PKM - Personal Knowledge Management

Looks like nothing really works for me. Notes are still scattered around services, but the main hub is Obsidian and a vault in Dropbox, synced to my iPhone with Syncthing, because... that's how it is.

Editors for iOS

1Writer

iA Writer

Drafts

Sync

Dropbox

Seems to be the most reliable, with the difference that it doesn't work with Obsidian on iOS :D

Remotely Save

https://github.com/remotely-save/remotely-save

Seems to be abandoned: https://github.com/remotely-save/remotely-save/issues/1087

Obsidian Live Sync

Alternativ Sync Server for Obsidian [[Mac Mini Homeserver#Obsidian Live Sync|See Setup in CasaOS]]

Syncthing

There are two Syncthing clients on iOS

I slowly move some vaults out of Dropbox and use Syncthing on iOS (via MoebiusSync).

Obsidian vaults are stored in the Obsidian iOS Sandbox, Möbius Sync can access these vaults and syncs them to the [[Mac Mini Homeserver#Syncthing|NAS]] or any other machine and vice versa. There are some problems with syncing Obsidian settings, for example hidden folders can't be accessed on iOS, but there are workarounds with apps, see the following links.

The sync works well so far (2025-05-26). It's not as convenient, but this is the only way on iOS besides iCloud and ObsidianSync.

Systems

Here are some systems that don't work for me, because I'm too busy reordering my notes and do all the meta work. Eventually I'll develop my own system and I promise, I wont bother anyone with a write-up why this is the best thing ever... #WIP

Pillars, Pipelines and Vaults (PPV), August Bradley

https://esilva.net/tla_insights/ppv_bradley

PARA

https://klemet.github.io/Workshop-Organization-EN/08-PARA.html

A Case Against AI Driven Second Brain Slop

by Joan Westenberg

[0:00] In 2007, a team of researchers at the University of Minnesota ran a study on choice paralysis. They gave one group of shoppers six varieties of jam to sample, and another group 24. The group with 24 options bought jam at roughly 1/10 the rate of the group with six. More choices, fewer decisions, more abundance, less action. The study's been cited thousands of times, mostly by marketers trying to sell you a curated subscription box. But the real lesson landed somewhere else entirely, and it took almost 20 years for it to show up.

[0:24] Somewhere in 2025, a critical mass of people hit a configuration sweet spot. They had Claude or ChatGPT hooked into their note-taking systems, their calendars, their reading lists. They could say, "Summarize this paper and file it under research and AI governance," and it would happen. They could say, "Generate a competitive brief on company X," and 45 seconds later, a 1,500-word document would appear in the right folder with the right tags. They could say, "Connect this to my notes on platform economics," and backlinks would populate like mushrooms after rain. It felt like having a team, a research assistant, a librarian, a junior analyst, all working in parallel, all filing things exactly where you told them to. The problem is that people stopped reading the files.

[0:57] Scroll through any Obsidian forum, any Notion subreddit, any productivity community, and the same confession surfaces over and over. "I have thousands of notes that I can't find anything." The vaults are, by any reasonable measure, impressive, but they're also, by any honest measure, useless about 80% of the time.

[1:11] There's a concept in organizational theory called information overload, coined by Alvin Toffler in Future Shock in 1970. Toffler was worried about television, paperback books, the sheer volume of print media flooding post-war households. His concern was that humans have a cognitive bottleneck. We can only process so much before we start dropping things, skimming, defaulting to pattern matching instead of actual comprehension. Toffler was right, but he was thinking too small. He imagined overload as a river rising. What we have now is a firehose pointed directly at your face, and you are the one who turned it on.

[1:36] The generative AI tools available in 2026 can produce text and diagrams and structured data and slide decks and wiki pages and linked knowledge graphs and formatted reports faster than any human can read them. A single Claude session can output 10,000 words of organized, tagged, cross-referenced content in under 5 minutes. That's roughly the length of a short novella every 5 minutes, on demand. And people are saving all of it.

[1:55] Think about what happens when a company hires too fast. You go from 10 people to 50 in a quarter. Everyone's producing work, memos and specs and strategy docs and Slack threads and Notion pages and Confluence wikis. The output looks great. The org chart fills in. The shared drive balloons. But nobody can find anything. Nobody knows what decisions already made, what research is already done, what conclusions someone on another team already reached. The company starts to drown in its own productivity. This happens at startups constantly. The failure mode is always the same: too many documents, not enough readers.

[2:24] Generative AI turned every individual into that over-hired company. You now have employees, agents, automations, prompt chains, dropping documents into various folders. And you can't be across every document. You're either the CEO of a 50-person org where all 50 people report to you, and they're tireless, and they don't need lunch breaks, they produce clean copy, and they never once stopped to ask, "Does anyone actually need this?" The documents keep arriving, the folders keep filling, the backlinks keep multiplying, and you keep telling yourself you'll review it all later. But you won't.

[2:48] In 2014, a blogger named Christian Tietze wrote about something he called the collector's fallacy. His argument was simple. Collecting information feels like learning, but it isn't. Saving an article to Instapaper feels like reading it, but it isn't. Clipping a passage to Obsidian feels like understanding it, but it isn't. Filing a PDF in your reference manager feels like having absorbed the argument, but you haven't. The gap between collecting and knowing has always existed. What's changed is that the collection side has now been almost entirely automated. You don't even have to do the collecting anymore. You can tell an AI to research X and save your findings, and the collecting happens without you lifting a finger.

[3:18] Try running an audit on your own vault or note system. How many have you opened in the past 90 days? And of those, how many did you actually read, not just glance at? And of those, how many changed how you thought about something or directly informed a decision? I think for most people, the hit rate lands somewhere around 1 to 5%. And the rest is dead weight with good metadata.

[3:37] The disk space is basically free. That's not the problem. The problem is cognitive. Every note in your vault is a small claim on your attention. And it's not because you're reading it, it's because you know it's there. You know there's a research brief on AI regulation you haven't reviewed. You know there's a competitor analysis from 3 weeks ago that might be outdated. You know there are 14 meeting summaries you told yourself you'd synthesize into your quarterly review. This ambient awareness of unprocessed information creates a low-grade anxiety that psychologists have studied for decades.

[3:58] In 2007, a team at the University of London found that the cognitive penalty of having unread emails sitting in your inbox was comparable to losing 10 IQ points. The emails didn't have to be urgent, they just had to exist unprocessed in your peripheral awareness. Now scale that from an email inbox to an entire knowledge management system that's being fed by tireless AI agents 24 hours a day. The vault isn't a second inbox — it never stops filling.

[4:19] A product designer sets up Claude to automatically generate a design critique for every screenshot she saves to a specific folder. Smart, right? Automated feedback. Except after 2 weeks, she has 340 design critiques she hasn't read. And the sight of the folder gives her a stomach ache. A consulting firm sets up AI-generated client briefing docs. Before every call, an agent pulls the client's recent news and financial filings and social media activity and industry trends into a three-page summary. The summaries are good, but also never read, because the consultants are already on back-to-back calls and have six other briefings waiting. A writer sets up an automation to generate weekly connection maps between all the notes added to his vault that week. The maps are beautiful. Obsidian's graph view lights up like a constellation. He shows it to people at parties. He never once uses the notes to actually write anything.

[4:58] The generation is easy. The consumption is the bottleneck. And the bottleneck isn't a technical problem, it's a human problem. The instinct is to organize better. More folders, better tags, smarter search, AI-powered retrieval so you can query your own vault like a database. And this helps a little, but it's treating the symptom. The disease is overproduction. You don't need a better filing system for 4,000 notes, you need to not have 4,000 notes. Or at least need to be honest about which of those notes are load-bearing and which are decoration.

[5:22] It's worth asking before telling AI to generate anything, "Will I read this within 48 hours?" If the answer is no, don't generate it. If you need it later, you can generate it later. The information isn't going anywhere. Claude's ability to produce a research brief on Tuesday is identical to its ability to produce one on Friday. The brief doesn't need to exist in your vault on Tuesday if you won't look at it until Friday.

[5:41] Manufacturing figured this out decades ago. Toyota's production system, developed by Taiichi Ohno in the 1950s, was built around a core insight: inventory is waste. Having parts sitting in a warehouse is a liability. It costs money to store, money to track, money to insure, and it becomes obsolete while it sits there. The just-in-time principle says, "Produce what you need, when you need it, in the quantity you need it." Too many people using generative AI are doing the opposite. They're running a just-in-case knowledge factory. Generate the brief just in case you need it. Summarize the article just in case it's relevant. Map the connections just in case a pattern emerges. File everything just in case future you wants it. But future you has never once been grateful for a speculative research brief.

[6:13] The people who have gotten a handle on this tend to converge on the same set of habits. They use AI in session for the thing they're working on right now. They ask questions, get answers, use the answers, and let the conversation disappear. If they need to save something, they save it deliberately, with a reason attached. "This is the research for the blog post I'm writing today. This is the financial model I need for Thursday's board meeting." They delete aggressively. They treat their knowledge base like a garden, not an attic. Things that haven't been touched in 90 days get composted. The anxiety of deleting lasts about 20 minutes. The relief lasts weeks.

[6:40] The vaults get smaller. The hit rates go up. The stomach aches go away. Every document an AI agent produces carries a hidden cost. Someone has to read it, or decide not to read it, or feel guilty about not reading it. That cost is invisible in the moment of generation. It shows up later as a vague sense of being behind, of having a system that's smarter than you are, of owning a library you can't keep up with. And the tools keep getting faster. The generation keeps getting cheaper. The agents keep getting more capable of producing more things in more formats, with more connections to more of your existing work. And the constraint was never going to be generation. It was always going to be attention. And attention doesn't scale.

• #knowledge, #iOS, #editors, #sync, #ai, #second-brain
Marcus Obst
Güterweg 89b
09474 Crottendorf
Germany

+49 37344 133407
info@marcus-obst.de