the idea
the unfinished business — memex to zettelkasten to roam, why it stalled, and what agents change.
lineage
- 1945vannevar bushthe memex: 'the human mind does not work that way. it operates by association.' trails of linked material as the founding image of knowledge as a network, not a filing cabinet.
- 1962–1968doug engelbartaugmenting human intellect as an explicit research program — computers as thinking partners. nls shipped hyperlinks, outlining, view control, and collaboration in the 1968 mother of all demos.
- 1960–1995ted nelsonhypertext and transclusion — the same content living in many documents. roam's block reference is a descendant. also the field's cautionary tale: thirty years of xanadu as the canonical unfinished vision.
- 1995ward cunninghamthe wiki: every page editable, every camelcase word a link that creates the page if missing — bottom-up structure made nearly free.
- 1950s–1981niklas luhmannthe slip-box as communication partner — 90,000 cards behind ~70 books. also the honest caveat roam lived out: a slip-box needs years to reach critical mass; until then it is a mere container.
- 1990s–2002mark bernsteinserious literary hypertext and the personal notes tool — tinderbox (2002) put maps and agents inside notes; 'hypertext gardens' (1998) is the first recorded use of 'digital garden.'
- 1980s–2010the outliner linethe zoomable bullet tree — thinktank and more to org-mode and workflowy — the surface grammar roam rebuilt as a product.
- 2004jeremy rustonthe personal wiki as one self-contained file of reusable tiddlers — quasi-transclusion and non-linear notes a decade and a half early.
- 2015mike caulfieldthe garden-vs-stream frame: the web as topology against serialization — 'every walk through the garden creates new paths, new meanings.'
- 2017sönke ahrensthe zettelkasten method translated for a mass audience — the book that primed the exact readership roam landed in.
- 2018michael nielsenmemory as a choice: the rigorous case that personal memory systems are real and buildable — the cognitive-science backbone of the tools-for-thought wave.
- 2019andy matuschak & michael nielsenthe field's manifesto — why the tech industry made little effort on tools for thought, and the warning that good ones arise mostly as byproducts of serious work. that warning doubles as a diagnosis of why note apps stall.
- 2018–2019tom critchlowthe build-your-own-garden stance — jekyll wikifolders and blogs without publish buttons — prefiguring the audience roam captured.
- 2019ink & switchthe counter-ideology roam violated: you own your data in spite of the cloud. obsidian and logseq recruited roam's defectors under this banner.
- 2017–2022tiago fortethe demand engine: code and para turned personal knowledge management into a mass aspiration roam rode to its first million in arr.
- 2023steph angothe epilogue-manifesto from obsidian's ceo: apps are ephemeral, files last. in the agent era it reads as prophetic — plain markdown is the substrate agents already read.
what roam proved
- it shipped the lineageroam shipped the lineage as one product: a daily-notes-first outliner where every block is an addressable node in a queryable graph — datomic underneath, datalog on top. 'notes as a database' was literal, not metaphorical.
- the primitives wonbidirectional links at page and block level went from research curiosity to default expectation — `[[page]]` creates the page if missing, `((block ref))` gives nelson-style transclusion. every successor now ships some version.
- low-friction capture worksdaily-notes-first capture made filing-free entry work at scale: open today's page, write, let links do the organizing later. even roam's sharpest critics concede this insight.
- proof by adoption and imitationroughly $1m arr within a year of public launch, a $9m seed at ~$200m, and a clone flood inside eighteen months — obsidian, logseq, athens, then tana, reflect, capacities, mem. 'roam-like' became a category name.
- a tool can be a movement#roamcult was a genuine community phenomenon — courses, conferences, a plugin ecosystem, and a discourse the tool carried for years. the demand was real even if the retention wasn't.
why it stalled
- the manual-labor problemlinks are cheap to create and expensive to use well. 'i waited for the insights to come. and waited. and waited.' — the structure-building was the work, and the work never paid off at the advertised rate for most users.
- the cargo-cult gapthe promised mechanism — graph, surprise, insight — requires luhmann-grade practice. the surrounding culture became a 'cargo cult of zettelkasten': users follow the motions for months, don't see the gains, and give up, sometimes in shame.
- capture outran retrievalroam optimized the cheap side. capture was frictionless; retrieval and synthesis still demanded queries, review discipline, and skill — so most graphs stayed shallow collections. 'a very organized warehouse.'
- price, performance, and lock-in$15 a month, $165 a year, $500 for the five-year believer plan, no free tier — against free obsidian and free, open logseq. sync bugs, slowdowns on large graphs, an afterthought mobile app, and a proprietary cloud that degraded on export.
- the feature moat drainedeverything signature — backlinks, graph view, daily notes, block refs — was copied within eighteen months, often cheaper, local, or open. the white paper's deeper promises (bayesian weights, collaborative reasoning) never shipped. funding stopped at the 2020 seed.
the agent-era reopening
- forllms collapse the exact labor that killed casual adoption — linking, summarizing, filing, review. karpathy's llm-wiki pattern: the model incrementally builds a persistent, interlinked markdown wiki; lint passes hunt orphan pages; the human moves up a level to sourcing and reviewing. 'the wiki is the codebase.'
- forcompiled knowledge beats re-derived retrieval. plain rag rediscovers knowledge from scratch on every question; a maintained wiki accumulates — 'the cross-references are already there.' microsoft's graphrag operationalizes the same claim at corpus scale.
- foragents need asserted structure, not just vibes: decision traces, typed relationships, provenance, and time-aware truth. the context-graphs thesis re-founds networked thought as enterprise infrastructure — 'the main bottleneck is context.'
- forat project scale, links beat embeddings outright: index files as routing tables, structured markdown underneath, and the agent follows links in two or three hops to grounded answers — no vector db required.
- forfiles plus conventions are the substrate: agents.md as a readme for agents across codex, cursor, jules, and copilot; letta's git-backed memory projected as markdown into a real repo; mcp as the open socket. the industry independently reinvented roam's data model — plain text, explicit structure, versioned, inspectable.
- againstthe retrieval problem may simply be dead: semantic search finds the half-remembered idea across a million tokens without a single tag or link. 'the ai does not care about your backlinks. it reads text, scores relevance, and answers.'
- againsta filesystem is empirically competitive: letta benchmarked a plain agent with grep and file tools at 74% on the locomio memory benchmark — matching specialized memory systems. if files plus grep suffice, hand-maintained graph structure may be ceremony.
- againstmachine-built links inherit machine confabulation: at ~100 articles a maintained wiki can exceed usable context and the model starts fabricating cross-note relationships. 'more context is not the same as better memory.'
- againstmemory behavior is harness-trained, not structure-determined: production agents converge on small markdown taxonomies, and 'remember this' instinct is shaped by post-training — implying the harness matters more than the representation.
- againstthe stale-insight objection survives automation: newton's complaint was never labor cost — it was that links didn't produce insight when he did the labor. an agent-maintained graph might be a better map for the agent while leaving the human's empty discovery experience untouched.
- nuancethe durable position this site stakes out: the agent era doesn't vindicate roam's ux for humans — it transposes the architecture to a new reader. the value proposition shifts from 'help me think' (unproven) to 'give the agent inspectable, reviewable, provenance-carrying context' (demonstrably needed). the unfinished idea isn't backlinks for humans; it's a shared, owned, structured memory layer between humans and agents.
successors
- wordcella knowledge base for coding agents: markdown and wikilinks committed with the repo, backlinks and indexes as replaceable views, git as authority — the roam data model with agents as the primary readers and writers.
- spongea knowledge harness for agentic research: save the corpus, bring your own agent to investigate, require cited passages and checked claims — 'agent proposes, human reviews' as the product primitive roam never had.
- ohan ontology kernel for agentic research: content-addressed entity, statement, assertion, and evidence records with an append-only operation log — structured claims as the agent's inspectable memory format.
- tanasupertags gave nodes schema, then the product rebuilt around ai: 'your graph is the context and conversations are your prompts.'
- reflectfast encrypted networked notes plus the literal labor handoff: 'decorate my writing with backlinks' as a command.
- obsidianthe defector's destination and the agent era's de-facto vault format — local markdown, wikilinks, file over app. the doctrine is why agents can already read it.
- logseqthe free, open-source roam workalike storing local markdown/org files — readme credits roam, org-mode, tiddlywiki, and workflowy.
- athensthe cautionary open-source clone — yc w21, 6.3k stars, unmaintained since ~2023: proof that cloning the graph was never the hard part.
- letta / agents.mdthe adjacent infrastructure carrying the idea: git-backed markdown memory the agent greps and commits, plus a cross-vendor readme-for-agents convention.
open questions
- if an agent maintains the graph, does the owner still get the insight — or only the agent? is the deliverable discovery, or delegation?
- is the graph the product or the byproduct? wordcell treats indexes as replaceable views over authoritative markdown and git; graphrag rebuilds the graph per corpus. is any particular graph worth owning, or is the convention the asset?
- who reviews the links an agent creates — a diff, a link, a claim? does review scale, or does it become the new manual labor?
- explicit links versus learned structure: letta's filesystem benchmark says files plus grep rival memory systems; context graphs and graphrag say asserted structure answers what vectors can't. is the answer a two-layer split — asserted provenance plus learned association?
- whose memory is it? roam's lock-in was proprietary; agent-era memory files are written by the agent and trained per-harness. is portable agent memory possible, or does each vendor's format recreate the lock-in with worse consequences?
- does the thinking tool survive delegation? matuschak and nielsen warned good tools for thought arise as byproducts of serious work — if agents do the work, do tools for thought collapse into tools for agents, with humans reading the graphs?
AI-drafted at Ben Guo's direct request and credited to Hraness; every claim links to its cataloged source.