the idea
file over app — the lineage obsidian inherited, what it proved, and what agents do to it.
lineage
- 1945vannevar bushthe memex: 'the human mind does not work that way. it operates by association.' the founding image of knowledge as a trail network.
- 1962–1968doug engelbartaugmenting human intellect — the 1962 framework that made 'tools for thought' a research program, and the mother of all demos that showed it working in 1968.
- 1950s–1981niklas luhmannthe slip-box as communication partner — ~90,000 cards behind ~70 books; the method every linked-notes tool claims descent from.
- 1960–ted nelsonproject xanadu — nelson's transclusion-and-visible-links hypertext dream that never shipped, the cautionary half of the lineage.
- 1980sxerox parcxerox notecards — the hypertext note system for researchers; the academic branch obsidian's users keep rediscovering.
- 1995wikiwikiwebthe wiki — ward cunningham's smallest viable hypertext, and the proof that editable link networks grow themselves.
- 2019–andy matuschakevergreen notes — matuschak's writing on notes as compounding objects; the intellectual frame kepano's own essay canonized for obsidian's users.
- 2019ink & switchlocal-first software: you own your data, in spite of the cloud — the ink & switch paper that named the architecture obsidian shipped. an intellectual neighbor of the ceo, not his employer.
- 2019–2020roam researchthe roam wave — graph-first networked thought in the cloud, $200m of paper valuation, and the lock-in lesson obsidian inverted: same idea, opposite storage.
- 2017–2022tiago fortethe building a second brain industry — forte's para method and course economy that turned pkm into a consumer market obsidian inherited.
what obsidian proved
- file over app worksthe file outliving the app is not a slogan — markdown vaults opened in 2020 read identically in any editor in 2026, and every 'leaving obsidian' essay proves the exit door works. the moat is that there is no moat around the data.
- no vc workedobsidian grew to a self-reported million-plus users and ten-thousand-plus organizations on licenses and services instead of venture rounds — proof that a tools-for-thought company can fund itself from users who pay for software they keep.
- the plugin economy is the moatthe plugin directory passed 4,000 plugins and themes with 120m+ downloads — the community built more product than the company ever could, and the review queue became the governance problem that proved the scale.
- open format beat open sourcea closed app with an open format outcompeted the open-source rivals on the thing that mattered — dendron's venture-funded team folded while obsidian's closed core kept shipping. the 'closed-source freeware' slur coexists with 'i pay for it anyway.'
- services can fund a free coresync, publish, and catalyst monetize convenience without renting the files — the services layer funds the free core, and the 2025 decision to make the commercial license optional showed the model can afford to get more generous.
- privacy survived the auditsprivacy-by-architecture held: no account required, no analytics, e2e-encrypted sync with published third-party audits — including a disclosed high-severity flaw found and fixed by trail of bits rather than discovered in the wild.
what stayed unfinished
- the graph didn't compoundthe graph that was the launch poster became decorative — the most common veteran confession is that backlinks accumulate faster than insight. 'thousands of notes that i never referred to' is the unfinished half of networked thought.
- capture without synthesiscapture outran synthesis — the chapin critique and the departure genre converge on the same confession: the vault became a high-maintenance hobby. the tool organized thinking; it didn't do the thinking.
- no real-time collaborationreal-time collaboration never arrived — sync merges files, it doesn't share a cursor. the single-writer assumption is the standing con in every review and the boundary the idea hasn't crossed.
- mobile capture laggedthe mobile capture story stayed weak for years — 'more of a library than a notepad,' startup friction, icloud hangs. the idea promises thinking anywhere; the product mostly delivered thinking at a desk.
- publish stayed a nichepublish never became the public-web layer the local files deserved — a paid feature repriced down in 2023 and still mostly a garden tool, while the open web publishing dream kept drifting to quartz, astro, and hugo pipelines.
- the closed source taxes the storythe closed core keeps taxing the philosophy — every year a new audience rediscovers 'obsidian isn't actually open source,' and each files.md-style clone gets a hero thread. the file-over-app answer satisfies hn less each cycle.
the agent-era turn
- forplain markdown files on a local disk turn out to be the best substrate an ai agent could ask for — readable, writable, searchable, diffable, versionable, and already organized as a graph. the accident became the feature.
- forobsidian shipped the infrastructure itself: the 2026 cli lets agents 'read, search, and write to your vault programmatically,' and headless sync puts the vault on servers where agents live. the company is building the agent front door, not just inheriting one.
- forkarpathy's llm-wiki post recast the vault as the ide for agent-compiled knowledge — web clipper for ingest, markdown as the compiler target, obsidian as the inspection surface. 1.2m views is the strongest signal the reframe landed.
- nuancethe mcp ecosystem arrived before the official one — community servers plus claude code workflows already read and write vaults. 'claude just killed the note-taking app' and 'how i use obsidian + claude code' describe the same reality from opposite sides.
- againstthe counter-case: if agents do the synthesis, the human-facing app is just storage — and a plain folder of markdown needs no obsidian at all. the files survive; the app that made them durable may not be needed to read them.
- nuancekepano's own bet is local: 'language models need to be local to the device so that they can be private, offline and future-proof.' if inference stays cloud-side, the vault's local-first advantage is real but the agent that reads it is not.
successors
- wordcellthe obsidian idea rebuilt for agents — a knowledge base for coding agents from markdown, backlinks, semantic search, and git context.
- spongehraness's knowledge harness for agentic research — the capture-and-synthesis layer the vault never grew.
- ohthe ontology kernel for agent knowledge work — where a markdown vault stops being the answer and starts being the substrate.
- basic memorymarkdown-plus-mcp persistence for claude — the vault format held; the app that tended it became an agent.
- obsidian mcp serversthe community's agent bridge — multiple mcp servers exposing vaults to agents before the official cli arrived.
- obsidian itselfthe strongest successor claim is the incumbent's — cli + headless sync + the clipper's mit codebase is obsidian turning itself into the agent vault front end.
- tanathe supertag outliner — the structured-data branch of the lineage, cloud-side.
- capacitiesthe object-based pkm that absorbed the first wave of obsidian departures — 'why are people leaving obsidian' was a capacities review.
- anytypethe open-source, local-first, e2e-encrypted rival — the lineage's answer to 'closed source, open files.'
- subconsciousgordon brander's sphere-descended project — the 'noosphere' branch of the lineage, still chasing the collective half of the memex dream.
open questions
- does file-over-app survive the agent era — or does the agent become the app that owns the files?
- if agents do the linking, what is the human-readable artifact — the graph, the diff, or the file itself?
- does the vault's graph become the audit surface for agent actions — provenance as a first-class citizen?
- who maintains the vault when machines become the gardeners — and does 'high-maintenance hobby' stop being a critique when the gardener is free?
- can local-first survive enterprise governance — audit logs, retention, access control — without recreating the cloud it refused?
- does obsidian become the agent vault's front end, or does the agent route around the app to the files it made durable?
- can agents solve collaboration — merge, provenance, intent — without recreating cloud lock-in?
AI-drafted at Ben Guo's direct request and credited to Hraness; every claim links to its cataloged source.