The first version of our knowledge map looked like a real map. Nodes, edges, clusters. It was beautiful, it was technically impressive, and after six months of use we had to delete the feature.
Key takeaways
- Beautiful maps get abandoned. The original force-directed map lost 88% of its weekly openers by week 4; a quieter, path-preserving rebuild held 58%.
- Three primitives replaced the graph. Anchors (explained ideas), bridges (connections), and threads (the questions that built each anchor).
- One good edge beats a hundred nodes. 71% of tutor-offered seam cards were followed by a learner question in the same session — versus 19% for the old "related ideas" offers.
- Maps must age. Fading nodes show where recall is weakening; load-bearing anchors stay bright.
The map that did not survive
The map showed every concept a learner had touched. The edges were weighted by how often the concepts appeared together in sessions. The clusters were computed automatically from the co-occurrence graph. The visualisation was a force-directed layout that updated live.
Learners stopped opening it.
The post-session interviews were unsentimental. The map told them what they had asked about, but not what they had learned. It treated every node as equal, when in fact some nodes were anchors (a definition that everything else hung from) and some were waypoints (a passing curiosity that did not need to be remembered). And it had no sense of time—no record of how the learner had come to know the idea, which questions had opened it, which detours had made it stick.
The first version of the map was a graph in the strict sense: nodes for concepts, edges for relationships, laid out topologically so that prerequisites appeared left-to-right and applications appeared right-to-left. The graph was, on day one, beautiful. It showed the structure of the learner's topic at a glance. It was the kind of diagram a textbook author would have been proud of. It was also the kind of diagram the learner stopped opening after the third visit.
We learned the reason by watching what the learners actually did. They opened the map when they had a question they could not place. They looked for the answer. They closed the map. The next time they had an unplaced question, they opened the map, looked for the answer, and closed the map. The map, in their hands, was a lookup table. The relationships between the nodes were, to them, scenery. They never followed an edge from one node to another. They never used the map to navigate; they used it to locate. And once they had located the few nodes that mattered to them, the map became a less efficient way to get to those nodes than a search bar.
This is the failure mode we missed. We had built a map that was true, and we had assumed that truth was the goal. The data said otherwise. The learners were not using the map to learn the structure of the topic. They were using the map to find the one or two nodes that the structure of the topic implied they should look at next. A static diagram of all the relationships was overkill. A pointer to the next node was enough.
The deeper problem was that the map was not for the learner in any active sense. It was a representation of the topic, not a representation of the learner's relationship to the topic. It did not change as the learner changed. It did not reward returning. It was, in the language we have come to use internally, a map of the territory, not a map of the path.
What a useful map has to do
A useful map, we now believe, has to do three things that the original map did not.
First, it has to preserve the path. A node is not just a fact; it is the result of a sequence of questions, examples, and corrections. When the learner returns to the node, they should be able to see the path that brought them there—not in full, but in enough detail to remind them how the idea was assembled.
Second, it has to age. Some nodes should fade as the learner's recall weakens; others should stay bright because the idea is now load-bearing. The map should make it obvious which is which, so the learner can spend their time on the things that need attention.
Third, it has to be navigable in plain language. The original map made us feel clever. The current map should make the learner feel capable. The same information, drawn so it can be read at a glance.
The first rule—seven edges—is the most counterintuitive. It feels like an arbitrary cap, and in some sense it is. The number seven is borrowed from the cognitive-load literature on working memory, where seven (plus or minus two) is the rough limit on the number of items a learner can hold in mind at once. The argument is not that a learner can only think about seven things at once; it is that a learner who is looking at the map is doing so to find a next step, and a next step is a single decision. A map that shows seven edges is a map that offers seven possible next steps, which is the most a learner can hold in mind while making that decision. A map that shows seventy edges is a map that requires the learner to filter before they can decide, and the cost of the filtering is, in our data, larger than the benefit of the additional options.
The second rule—edges fade when unused—was the design choice that took the longest to justify. The intuition was simple: if a learner has not followed an edge in thirty days, the edge is not currently part of the learner's working model of the topic, and showing it at full weight misrepresents what the topic currently is for them. The empirical justification was harder. We had to show that fading did not cause learners to forget edges they would later need, and we had to show that fading did improve the rate at which learners opened the map at all. The data, summarised below, does show both, with the caveat that the four-week recall effect is not yet significant.
The third rule—surface the seam, not the network—is the rule that ties the map to the rest of the tutoring system. When the tutor offers a related-idea card, it does not pick the related idea at random from the learner's graph. It picks the related idea that connects to the learner's current question at exactly the place where the learner has been showing a seam. This is the same diagnostic we use for explanation episodes (see the note on "explanation evidence"), applied to the map. The map is no longer a passive display of relationships. It is an active reader of the learner's current state, and it offers the next step on the basis of that reading.
The point of a map is not to show you what is on the territory. The point is to help you decide where to walk next.
What we built instead
The current knowledge map is built around three primitives: anchors, bridges, and threads. Anchors are ideas the learner has explained cleanly in their own words; they stay bright. Bridges are ideas that connect two anchors; they stay visible but quieter. Threads are the sequences of questions that brought the learner to each anchor, and they are stored on the anchor itself.
The result is less impressive to look at and much more useful. When a learner returns to an anchor, they can see the thread that built it. When an anchor is fading, they can see which bridges depend on it. When the learner is starting something new, the map shows them which anchors are likely to be useful—not by topic, but by how well they remember them.
The most striking single observation from the rebuild is that learners stopped closing the map. In the old map, the median learner opened the map 1.2 times per week and closed it within 30 seconds. In the rebuilt map, the median learner opens the map 2.7 times per week and stays open an average of 4.1 minutes. The map has become a place the learner visits, not a thing the learner consults. We attribute most of the difference to the fade behaviour: the rebuilt map is less crowded, so the act of opening it is less expensive, and the edges it shows are the ones the learner has been using, so the act of opening it is more rewarding.
The second striking observation is that the rebuilt map produced edges the old map did not. In the old map, the edges were a fixed property of the topic. In the rebuilt map, edges are weighted by recent use, and a learner who has just spent a session on, say, related rates in calculus will see the edge between "related rates" and "implicit differentiation" at full weight, even if the topic graph did not originally weight that edge heavily. The map is, in this sense, learner-specific. Two learners studying the same topic can have visibly different maps, and that difference is, in itself, useful. The map is no longer a textbook figure; it is a portrait of the learner.
A third observation, more tentative: the rebuilt map seems to be a better predictor of what the learner will ask next. We have a small internal model that, given the current state of the map and the learner's last question, predicts the next question with about 41% top-one accuracy and 68% top-three accuracy. The same model trained on the old map produced 22% and 44% respectively. The map is now a real signal, not just a display.
Weekly map opens across the 12-week cohort
What we are still working on
Two open questions. The first is the question of scale: a map of fifty anchors is a pleasure; a map of five hundred is overwhelming. We have not yet found a way to make the large map feel as navigable as the small one, and the approaches we have tried—automatic clustering, time-windowing, importance scoring—each have a way of being wrong that the learner notices.
The second is the question of editing. Learners sometimes want to remove an anchor that they no longer care about, or to merge two anchors that turned out to be the same idea. The current map lets them do this, but the undo flow is awkward. We are still working on the right verb set for editing a learning map, and we will publish what we land on.
The phrase "a dashboard for a job I did not have" came up four times in the original cohort and zero times in the rebuilt cohort. We have come to use it internally as the canonical description of the failure mode we were trying to fix. A map that shows everything the learner could be doing is, in practice, a map that shows a job the learner does not have. The map was, for those learners, a list of obligations. The rebuilt map is, by contrast, a list of options. Obligations repel; options invite.
The point of a map is not to show you what is on the territory. The point is to help you decide where to walk next.
What the data says
From April to June 2026 we compared 96 active learners on the original map and the rebuilt map. The rebuilt map was not more popular in week one—it was more persistent:
Map cohorts · 96 learners, 12 weeks
Apr – Jun 2026We also measured whether the map changed what learners did next, not just what they looked at. When the tutor offered a seam card—a single edge where two ideas actually meet—71% of those offers were followed by a question in the same session. The original map, which showed everything, produced follow-up questions from only 19% of its "related ideas" offers.
Which interactions carry the map
Breaking the rebuilt map into its primitives shows the anchor card, not the seam card, is what learners return to — while the seam card is what converts looking into asking.
| Interaction | Offer → question | Revisit within 7 days |
|---|---|---|
| Seam card (a bridge edge between two anchors) | 71% | 64% |
| Anchor card (idea + its thread) | 58% | 71% |
| Original “related ideas” offer | 19% | 22% |
Table 1 — Offer-to-question and revisit rates by interaction type · n = 96 learners
How we measured it
Ninety-six active learners were assigned to the original or rebuilt map for twelve weeks. We logged weekly opens, edge follows, anchor revisits, and whether a tutor-offered edge was followed by a learner question. Eighteen post-session interviews explained the mechanism: learners described the rebuilt map as "quieter" and the original as "a dashboard for a job I did not have".
The eighteen interviews were semi-structured and lasted 25–40 minutes each. We asked learners what they used the map for, when they opened it, when they closed it, what they wished it would do that it did not, and what they wished it would not do that it did. The interviews were transcribed, anonymised, and coded by two reviewers independently. We did not measure whether the rebuilt map improved learning outcomes directly. We measured whether learners used the map more and whether the edges they followed led to more follow-up questions. The assumption is that more follow-up questions is, on average, a leading indicator of more learning, but we have not yet closed the loop with a delayed-recall measurement. That measurement is on the calendar for October 2026.
Limitations
The cohort is 96 self-selected learners from one product, split 48 per arm—wide confidence intervals on every percentage we report. The twelve-week window cannot say whether rebuilt-map use survives a semester. The interview sample (18) is large enough to explain a mechanism and too small to rank one. And because the rebuilt map shipped with other April changes, we cannot fully separate the map's effect from the release it rode in on.
Open questions
Scale remains unsolved: we do not yet know how the rebuilt map behaves past five hundred anchors. Editing needs a better undo flow. And the hardest question is whether map use improves delayed recall, or only session engagement—the numbers we have so far suggest both, but the recall effect is not yet significant at four weeks.
A fourth open question, which we have not yet studied: does the map work the same way for learners who are studying for an exam as it does for learners who are studying out of curiosity? Our intuition is that exam-driven learners want a more complete map, because they want to know what they are responsible for, and curiosity-driven learners want a more selective map, because they want to know what to look at next. The current rebuild is biased toward curiosity. We do not yet know whether the bias is a feature or a bug.
A fifth open question: how should the map behave when the learner switches topics? A learner who has spent three months on calculus and now switches to linear algebra has a calculus-heavy map and a linear-algebra-light map. Should the map carry the calculus edges with them, in case they return? Should the map drop the calculus edges, on the theory that they are no longer part of the learner's current model? Should the map offer both, side by side? We do not yet have a settled answer.
References
- Nesbit, J. C., & Adesope, O. O. (2006). Learning with concept and knowledge maps: A meta-analysis. Review of Educational Research, 76(3), 413–448.
- Novak, J. D., & Cañas, A. J. (2008). The theory underlying concept maps and how to construct them. Technical Report IHMC CmapTools.
- Kintsch, W. (1988). The role of knowledge in discourse comprehension: A construction-integration model. Psychological Review, 95(2), 163–182.
- Roediger, H. L., & Butler, A. C. (2011). The critical role of retrieval practice in long-term retention. Trends in Cognitive Sciences, 15(1), 20–27.