NoteCards, KMS, and the knowledge-base lineage
The third strand of pre-web hypermedia, after Xanadu’s publishing ambitions and Brown’s literary-academic ones, was the knowledge-base lineage. It treated hypertext primarily as a tool for thought. The systems in this strand were built by their researchers for their researchers to use as instruments of intellectual work — design, theorem proving, writing, software architecture, argumentation. The systems were used in production, by their builders, for years at a time, on real problems. They produced a body of practical knowledge about what hypertext was useful for, what it was not useful for, and what shape a hypertext system had to have to support sustained intellectual work. The two flagship systems in this strand are NoteCards, built at Xerox PARC by Frank Halasz, Randy Trigg, and Tom Moran in the mid-1980s, and KMS, built by Knowledge Systems Inc. as the commercial descendant of CMU’s ZOG project. A third project, gIBIS, sat alongside them and pioneered the argument-mapping tradition that has been the most directly recovered in the recent personal-knowledge-tools wave. None of these systems is famous outside the small history-of-hypertext literature. All of them are, in their substantive contributions, more directly present in the current generation of tools (Roam, Obsidian, Logseq, Tana) than any of the more-celebrated lineages.
NoteCards at PARC
Xerox PARC is the research lab whose work in the 1970s and 1980s shaped almost everything about modern personal computing, mostly indirectly through the people who left it. NoteCards was a PARC project of the mid-1980s, after the Alto and Smalltalk and the laser printer and Ethernet had already established PARC’s reputation, and at a moment when the lab was still funded well enough to produce serious research software. NoteCards was developed primarily by Frank Halasz, Randall Trigg, and Thomas Moran. Halasz was the technical lead; Trigg, who had completed a doctoral thesis at the University of Maryland in 1983 on a hypertext system called Textnet, brought a research history in linking-as-citation; Moran was a cognitive scientist who studied how people actually used the system. The combination — engineer, hypertext researcher, cognitive scientist — gave the project an unusual depth of attention to whether what the system did and what users did with it were the same thing.
The data model was simple. The unit was a “card”: a rectangular note holding some content (text, graphic, sketch, image, structured list, embedded subsystem) and located at a specific position in two-dimensional space inside a “filebox.” Fileboxes could contain other fileboxes, recursively. Cards could be linked to other cards by typed links: a link had a name describing the kind of relation it represented, and the type vocabulary was extensible by users. The link was a directed relationship; the back-link was automatically maintained so that, looking at any card, the user could see what other cards linked to it. The whole structure was browsable in multiple views: as a graph (showing cards as nodes and links as arcs), as a filebox tree, as a list of recently visited cards, as a search-result set.
What made NoteCards influential was less the data model — every hypermedia system of the period had some version of nodes and links — than the way users used it. Trigg, Moran, and colleagues published a series of empirical studies of NoteCards in practice. They watched researchers, themselves and others, use the system over long periods of time on real intellectual projects, and reported on what worked and what did not. The studies are still cited because they are unusually frank. The reports cover the use of NoteCards to write papers, to design software, to plan experiments, to manage research correspondence, to draft talks. They cover failures: where users got lost in the graph, where the filebox hierarchy became unmanageable, where the system’s tools for restructuring did not match the way users wanted to restructure. They cover successes: how the link-types vocabulary evolved through use, how certain users developed distinctive personal styles of organization, how the practice of using NoteCards changed the way users thought about their work.
One of the central findings, articulated in Halasz’s 1988 paper “Reflections on NoteCards: Seven Issues for the Next Generation of Hypermedia Systems,” was that the system needed support for what Halasz called “composite nodes”: cards that were assembled from other cards, presented as a unit, but retaining the structure of their components. The user-facing problem was that an idea expressed across several cards was harder to manage than the same idea expressed in one card, because the system did not give the user a way to treat the cluster as a single object. NoteCards did not solve this fully; Halasz’s paper was, in significant part, a call for next-generation systems to solve it. The current generation of knowledge tools — Roam in particular — has solved it through the device of block-level transclusion, in which any block on any page can be referenced and displayed inside any other page, retaining its identity and updateability. Halasz’s seven issues were a research agenda; the recovery has not addressed all seven, but several have been addressed in narrow form.
NoteCards ran on Xerox D-machines — Dolphins, Dorados, and similar Interlisp-based workstations — that were never sold to a mass market. The system was used inside PARC and at a handful of external sites that had compatible hardware. The user community was perhaps a few hundred people at peak. When the D-machine line was discontinued and PARC’s funding model changed in the late 1980s, NoteCards stopped being developed; the research findings outlived the system. Halasz left PARC for academia and later industry; Trigg continued at PARC on related work; Moran moved into design research more broadly. The lineage of NoteCards’ specific design choices runs through subsequent PARC projects (Aquanet, VIKI, others), into Marshall and Shipman’s “spatial hypertext” research that continued through the 1990s, and eventually into the commercial wave that began in the late 2010s with Roam.
ZOG and KMS
The other major knowledge-base system of the period had a different lineage and a different feel. ZOG was a Carnegie Mellon project that began in 1972 under Allen Newell, with George Robertson as the principal designer and contributions from Donald McCracken and Robert Akscyn among others. The system was built on the CMU mainframes of the early 1970s and was, in its initial form, a menu-driven information retrieval system: screens of structured text with selectable items at the bottom, leading to other screens. The design was austere and the operations were fast: a ZOG screen could be navigated in fractions of a second, and the system was designed around the principle that the user should never have to wait for the next screen to appear. ZOG was deployed in production at CMU and aboard the USS Carl Vinson aircraft carrier, where the Navy used it for crew training and reference. The Vinson deployment is worth pausing on: it is one of the few examples in the history of hypermedia of a system being used by a non-academic, non-research user population for years at a time as their primary information system.
Akscyn and McCracken left CMU in the early 1980s to found Knowledge Systems Inc. and commercialize the ZOG approach. The result was KMS, the Knowledge Management System, released in the mid-1980s and developed through the late 1980s and into the 1990s. KMS preserved ZOG’s central commitment to speed: the system was designed so that navigation was effectively instantaneous, with sub-second response times on every operation. The data model was a flat space of frames; every frame was identified by a name and contained structured content. Links between frames were inline annotations attached to specific items inside a frame. There was no separate addressing scheme for sub-frame elements; the unit of linking was the frame, and the within-frame structure was an authorial convention rather than a first-class object.
KMS was sold commercially to engineering and information-intensive enterprises. It was used at Boeing for technical documentation, at Westinghouse for plant management systems, at a number of other industrial sites. The KMS user base was perhaps tens of thousands across the customer organizations. The system’s design choices — fast navigation, simple data model, integration with existing institutional document workflows — were optimized for the working-day use of professionals, not for the literary or research uses the other systems aimed at.
Akscyn and McCracken’s 1988 paper “KMS: A Distributed Hypermedia System for Managing Knowledge in Organizations” argued, against the more elaborate research systems, that hypermedia’s value at scale came from being unfussy. A user opening KMS could navigate any frame in the corpus in seconds; the system was on every desk in the organization; the frames were authored by people in the organization for their own use and the use of their colleagues. The argument was that this kind of low-ambition, high-frequency use was where hypermedia would actually live, if it lived anywhere, and that the research systems’ elaborate features were not what made the difference. The argument was, in retrospect, prescient about what wiki software would later look like. Ward Cunningham, the original wiki author, has cited KMS as one of his influences.
KMS continued to be sold and supported into the mid-1990s, when the web began absorbing the use cases. Knowledge Systems Inc. continued to exist as a smaller business and the KMS user base, like much else in this book, gradually migrated to web-based systems. The Knowledge Systems lineage did not produce a direct successor product; what the system had to offer was largely subsumed into the wiki and intranet generation that followed.
IBIS and gIBIS
A parallel lineage, smaller in scope but historically important, came from the work of Horst Rittel and Werner Kunz on Issue-Based Information Systems (IBIS) in the 1970s. Rittel had developed, in his work on what he called “wicked problems” in design and policy, a method for capturing the structure of an argument as a graph of issues, positions, and arguments. An IBIS diagram laid out the issue under discussion as a node, the positions taken on the issue as connected nodes, and the arguments for and against each position as further connected nodes. The structure was navigable: one could trace from an issue through the positions and arguments to see how a discussion had developed. Rittel proposed the method for use in design teams, planning bodies, and other settings where the rationale behind decisions needed to be made visible and durable.
gIBIS, the graphical IBIS, was a software implementation built at MCC in Austin, Texas, in the late 1980s by Jeff Conklin and Michael Begeman. gIBIS turned Rittel’s paper-based method into a hypermedia system: users could create issue nodes, position nodes, and argument nodes; could link them according to the IBIS conventions; and could view the resulting graph and navigate around it. The system was used in design meetings at MCC and at a number of customer sites. Conklin and Begeman published several papers reporting on the use, and Conklin’s subsequent book Dialogue Mapping (2005) extended the IBIS approach into a wider audience of facilitation practitioners. The method, distinct from the software, continues to be taught in some design and facilitation programs.
The IBIS / gIBIS lineage is the most direct precursor of the current generation of “Tools for Thought” that explicitly position themselves as supports for structured thinking. Roam Research, Obsidian’s various plugin extensions for graph view, and the smaller specialty tools (Heptabase, Kinopio, Scrintal) all carry forward the basic IBIS impulse: that the graph of a thinker’s argument is itself an artifact worth producing and preserving. The tools have moved away from IBIS’s specific argument grammar — they do not, generally, require nodes to be classified as issues or positions or arguments — but the underlying conviction that the structure of thought is worth making visible and navigable is the same.
What these systems shared
The knowledge-base systems differed in many specifics — NoteCards was graphical and Lisp-based, KMS was textual and optimized for speed, gIBIS was structured around a specific argumentation grammar — but they shared a set of assumptions that distinguishes them from the other hypermedia traditions in this book.
They assumed the primary user was the same person as the primary author. The systems were not publishing tools; they were thinking tools, used by their users to organize their own ongoing intellectual work. The model audience was the person sitting at the workstation, working on a long-running problem, and accumulating notes, links, and structures as they went. The system’s job was to make that accumulation useful — to make the user’s own past work findable, traversable, and combinable with their current work.
They assumed the work was durable. A NoteCards user accumulated cards over months and years on a sustained research project; a KMS user added frames to a corporate corpus that was expected to last as long as the corporation did; a gIBIS user was producing argument structures that would be referred to in later design decisions. The systems were optimized for long-term accumulation rather than short-term publishing.
They assumed the user knew what they were doing or would learn. The systems were not designed to be approachable to casual users. NoteCards required learning the link types, the filebox conventions, the navigation idioms; KMS required learning the frame-and-name addressing scheme and the inline-link conventions; gIBIS required learning the IBIS grammar. The expected payoff for the investment was that the user, once trained, would be able to do intellectual work the system specifically supported.
They assumed thinking was structural. The whole point of these systems was that ideas had structure, that the structure was worth capturing, and that captured structure could be reused and extended. This is in tension with the more-recent doctrine that good user interfaces should let users do what they want without imposing models. The knowledge-base systems imposed models, deliberately, because the models were the value.
The intersection of these assumptions is approximately the user the current personal-knowledge-tools wave has rediscovered. The Roam user, the Obsidian power user, the Logseq devotee — these are people who think structure is worth capturing, who are working on long-running personal projects, who have invested time in learning a specific way of organizing notes, who treat their notes as a substrate for their thinking rather than as a publishing medium. The lineage runs straight back to NoteCards.
Why these systems did not become mainstream
The reason the knowledge-base systems did not become mass-market software is largely the reason every system in this book did not become mass-market software in the pre-web era: they ran on hardware most people did not have, in software environments that required institutional support, with learning curves that excluded casual users. NoteCards required a Xerox D-machine; KMS required a workstation-class machine and the KMS license; gIBIS required MCC’s environment and the IBIS training. None of these were tools you could pick up at the consumer-electronics store. The systems were designed for the workstation era; they did not have a natural path into the personal-computer era; they did not survive the transition.
The web, when it arrived, did not absorb these systems’ specific capabilities. The early web was designed for publishing, not for sustained personal intellectual work. Wikis, which began with Ward Cunningham’s WikiWikiWeb in March 1995, recovered some of what KMS had been doing — fast, low-ceremony, collaborative editing of a structured knowledge base — for groups rather than individuals. Wikis spread, became substantial knowledge bases at organizations and online, and culminated in Wikipedia, which is the largest knowledge-base hypertext in the world. But Wikipedia is a publication, not a thinking tool. It is a finished artifact intended to be read; it is not, for most users, a substrate they use to do their own intellectual work. The space that NoteCards and KMS had occupied — the individual or small-team thinking tool, with hypermedia structure, used over years on long-running projects — remained mostly unoccupied through the 2000s and most of the 2010s.
The recovery, in brief
The recovery is the subject of chapter twenty-four. Here it is enough to note that it is real, ongoing, and substantial. Roam Research’s launch in 2017 (general availability 2019) marked the beginning of a wave of personal knowledge management tools that revived, in browser-based or desktop form, much of what the knowledge-base lineage had been doing. Obsidian, Logseq, RemNote, Tana, Heptabase, Reflect, Capacities, and a half-dozen others have all entered the same space. The tools differ in details but share a common ancestry: they treat notes as a network of linked, addressable, often block-level units; they support bidirectional linking; they offer graph-view visualizations of the link structure; they are designed for long-term personal use rather than for publication.
Most of the developers of these tools are aware, to varying degrees, of the historical lineage. Roam’s name is a reference to Engelbart’s “ROAM” — the early NLS prototype was sometimes called that — and the founders, particularly Conor White-Sullivan, have cited Engelbart explicitly. Obsidian’s developers have cited Vannevar Bush. The discourse around the current wave is sometimes uncomfortable for people who lived through the previous waves, because the new tools sometimes present themselves as recent inventions when they are recoveries; but the recoveries are real, and the users of the new tools are doing, in 2026, much of what the NoteCards users were doing in 1986.
The lost intermediate decades
The thirty-year gap between NoteCards (and KMS, and gIBIS) and Roam is worth marking. For most of those thirty years, the kind of work these systems supported was harder to do than it had been in 1986. There were attempts to fill the space: Lotus Agenda (1988-1992), various outliners (More, Ecco Pro), Personal Brain (since 1998 in some form), Tinderbox (since 2002), various specialty knowledge tools. None of them caught fire. The mass market was on the web, the web was a publishing system, and the publishing system did not give users a place to keep their own thinking.
The cost of this thirty-year gap is hard to measure. It is the difference, summed across a generation of knowledge workers, between having a tool designed to support sustained thinking and not having one. Most of the work that should have happened in such tools happened instead in email, in Word documents, in spreadsheets, in scattered text files, in browser bookmarks — in formats that were never designed for accumulation, that did not support linking in any serious way, and that produced no durable substrate. The intermediate decades are a loss the recovery is now slowly making up for, but the work that was not done because the tools were not available is gone.
The lineage section of the book ends here. The next part shifts from hypermedia systems to the other side of what the pre-web internet was: the protocols and architectures that proposed answers to the question of how networked services should be structured, separately from how documents within those services should be linked. Gopher, the first chapter in Part III, is the cleanest case of the contest. It was released the same year as the web, was the web’s principal competitor for two years, was technically simpler in many respects, and lost the contest in a way that turned on a single institutional decision that the people who made it now wish they had made differently.