An intelligent companion for the self-taught reader
I grew up in Norway, where history education is a jumble of disconnected topics. Vikings one week, the Bronze Age the next, the French Revolution the third. By the time you finish school, you’ve heard all the names — Charlemagne, Constantinople, the Silk Road — but nothing connects to anything.
For most of my life, this didn’t bother me much. Then I went to Crete.
Walking around Heraklion, we came across a massive fortress — walls several meters thick. A poster explained it had been besieged by the Ottomans for thirty years while held by the Venetians. I had lived in Italy, visited Venice many times, but never knew Venice had colonies. Then: “Before the Venetians, it was part of the Byzantine Empire.” I just stood there. I could not tell you a single meaningful thing about the Byzantine Empire. Constantinople — does that have something to do with Rome? At what time? I was 42 years old.
I had read the big history books — Sapiens, Guns Germs and Steel — and they felt fascinating in the moment. But I retained almost nothing. Worse: I couldn’t tell what was controversial. When everything is new, you can’t see the choices the author is making. You don’t have the background knowledge to think critically.
History is different from other kinds of reading. In a philosophy book, there are key ideas to extract. In a history book, every sentence contains new information — names, dates, places, causes. You feel cognitively overloaded, stressed because you feel like you should be remembering everything, and three months later it’s all mush.
I was almost giving up on reading nonfiction. The gap between what I understood while reading and what I could reconstruct months later felt demoralizing.
Petrarca emerged from years of following hunches, not knowing what would work.
ListenNotes, Otter.ai voice notes while walking, Roam Research. Already experimenting with voice capture for learning six years before AI. The discovery: speaking about what you’re learning forces prioritization in a way that note-taking doesn’t.
A century per month from Charlemagne to the present. Minimal goal: for each century, know the map, major wars, a paragraph summary. Starting from literally zero. Inspired by the Italian liceo classico, where the curriculum moves synchronously across art, music, literature, philosophy, political history. I got so interested in Charlemagne that I stayed with him for several months and never followed through to the other centuries. But it was an early attempt at the systematic scaffold I was looking for.
If the 12 Centuries project was wide and thin — scanning across centuries — this was the opposite: zoomed into a single place and a single year, reading Norwegian newspapers from 1814 day by day through 2025. History through primary source immersion.
A single-user Arabic learning app. Sentence-first learning, spaced repetition, capturing from real sources. ~1,300 words in 8 weeks. The crucial lesson: spaced repetition never worked for me in isolation. But the moment it integrated with real reading workflows, it became powerful.
The thread: voice capture works. Bounded structure works. Integration with real workflows works. AI changes what’s possible. Petrarca combines them — but the approach is open-ended and exploratory. I have a hunch about what a learning companion should do, and a lot of experimentation to find out.
What does it actually feel like to use this?
You started Norwich’s Sicily last week. It synced from Kindle automatically and the system mapped it against your curricula. You read three chapters on the couch last night. While reading chapter 3, you tapped the margin and spoke a quick voice note about something Norwich said about Archimedes. In chapter 4, you photographed a page with a map of Arab trade routes — the app found the page number and OCR’d the text. No interruptions to the reading itself. The book is the book.
Getting in the car, you see voice prompts are available. “What do you know about Frederick II Stupor Mundi?” No hints, no description — that’s the point: you have to recall it. You press record and talk for two minutes while driving. You mention his court in Palermo, the six languages, the Sixth Crusade. You wonder aloud: “Was he influenced by Islamic philosophy?”
78% coverage. Twelve facts captured, three missed. Two wonderings queued for the microlearning pipeline. By tomorrow, “Was Frederick influenced by Islamic philosophy?” will be a researched card. The missed facts get boosted priority in review. The things you forgot are exactly what you’ll see next.
A review card asks about the Siege of Syracuse. The temporal hook helps: Archimedes and Hannibal on opposite sides of Rome’s two-front war. You rate it “Hard.” Next: a microlearning card triggered by your wondering this morning — Frederick II and Islamic philosophy, with primary sources and places you can visit. “Palermo after Frederick’s death” — another wondering — is already queued.
Three nodes moved from “mentioned” to “engaged.” A new cross-curriculum connection: your voice recall about Frederick linked to the Arab Sicily chapter note, and the Arabic-Latin translations connected to Aristotle in Ancient Greece. The framing is always positive: here’s what you know, here’s how it’s growing.
Eight minutes of your day. Tonight: three more chapters on the couch.
Think of the Civilization game: the map is foggy except for the small area you’ve explored. You pick a spot and go deep — not a quick Wikipedia skim but multiple sources, historical fiction, documentaries, maps. You try to understand the people, the architecture, the living conditions. And from that patch of dry land, you expand to adjacent territory.
I described this in a short video using Charlemagne as my anchor. I started around 800 AD — the coronation as the new emperor of the Western Roman Empire. I ended up reading three historical novels in German, richly researched, covering the chronology but also the role of the church, the Germanic tribes, the pagan faiths. Then I could expand in every direction:
West to Spain — overrun by the Muslims, Charlemagne’s failed expedition to Barcelona, the encounter with the Basques that became the Song of Roland. East to Constantinople — Charlemagne’s complex relationship with Empress Irene, the worry that his coronation would be seen as claiming the whole Roman Empire. Further east to Baghdad — the Caliph who sent Charlemagne a white elephant that made it all the way to Germany and became a battle elephant (they found its bones in a river). South to Italy, the Pope, the Langobards. North to the Vikings — Alcuin, the great scholar at Charlemagne’s court, came from Northumbria and wrote about the attack on Lindisfarne. Backward: the fall of the Western Roman Empire, the migrating peoples. Forward: the split into Germany and France, the beginning of the Viking Age.
The same could be said about Caesar, Alexander, the Persian Wars. Each anchor opens adjacent territory. The knowledge grows outward like the explored area on a game map.
The anchors give you gist — deep, vivid, connected understanding. This is the most durable kind of historical knowledge. But specific facts matter too. Dates, names, sequences are the skeleton that everything hangs on.
An example. I picked up a Norwegian high school textbook on the French Revolution. It said people resented the church’s taxes. And I thought: only a hundred years earlier, France had a devastating civil war over the conflict between Protestants and Catholics. Tens of thousands killed. Not mentioned by word in that chapter. Does it matter? I don’t know. But I couldn’t have asked that question without knowing about the Wars of Religion. The factual scaffold made the critical question possible.
Sam Wineburg’s landmark study showed this concretely. Two historians read the same Lincoln documents. The specialist — who knew the sequence of days within specific months — defined the interpretive issues in four minutes. The non-specialist spent forty minutes “cognitively flailing” before reaching a narrower interpretation. Same intelligence, same skills. The difference was the facts. Facts are the speed multiplier for deep thinking.
Michael Nielsen makes the same point from a different direction. Drawing on chess research: experts don’t think differently, they chunk differently. A master recognizes 25,000–100,000 patterns as single units — expanded working memory. “Having more chunks memorized in some domain is somewhat like an effective boost to a person’s IQ in that domain.” If “the Punic Wars” is a rich chunk rather than an unfamiliar term, you can hold more of the argument in your head.
E.D. Hirsch extends this into a full theory of literacy: reading comprehension depends on specific background knowledge, not transferable reading skills. A striking study by Recht and Leslie gave students a passage about baseball: poor readers who knew baseball outperformed strong readers who didn’t. Domain knowledge trumped reading ability. Hirsch’s most dramatic evidence comes from France, where a 1989 shift from knowledge-based to skills-based curricula caused reading scores to decline and inequality to increase over twenty years. The knowledge was the infrastructure. Remove it and comprehension collapses.
This connects to cognitive load theory (John Sweller). Working memory holds roughly 7±2 items (George Miller). But “items” can be large, richly-encoded chunks or small, disconnected fragments. Prior knowledge reduces cognitive load by letting you process larger units. When “the Investiture Controversy” is a rich chunk, I can hold the whole argument in working memory. Without it, each unfamiliar name consumes a slot and the argument overflows.
So hooks and facts form a virtuous cycle. Deep anchors give you gist. The factual scaffold gives you the skeleton. Each supports the other. Petrarca maintains both.
From building Hamarquizen and my own experience, I’ve found four types of temporal connection, roughly ordered by how well they stick: (1) Anchoring to events I already know (“Archimedes died ~75 years after Alexander”). (2) Same-moment connections (“While Archimedes defended Syracuse, Hannibal was in Italy”). (3) Causal chains (“Archimedes’ death was a consequence of Hannibal’s invasion”). (4) Cross-domain surprises — but only if I already know the other domain. A connection to something unfamiliar isn’t a hook, it’s noise — so the system checks what I know before generating cross-domain links.
The core challenge: map the knowledge I have, the interests I’m pursuing, and the gaps I haven’t filled. History is impossibly fractal — you can zoom in forever. The system needs some way of answering: what are the important things to know?
Topic maps of 50–80 concepts per domain, generated by Claude Opus. Ancient Greece 67 nodes, Rome 55, Sicily 70. Roughly what a good lecturer would say you should understand. Bounded scope creates achievable goals and visible progress. There’s a tension: I might want to go deep on Frederick II or the Arab-Norman synthesis beyond any curriculum. The entity system and microlearning capture that depth. This balance will need revisiting.
Archimedes exists once in the shared entity database, but the Sicily curriculum sees “Archimedes the defender of Syracuse” while History of Science would see “Archimedes the mathematician.” Same entity, different knowledge states across curricula. When a node lights up in multiple curricula — a nexus point — those are the richest learning moments.
Unknown → mentioned → engaged → anchored. Always positive: “here’s what you already know.” Scheduling models decay underneath, but the visible level never drops.
Tracks books from Kindle (automatic sync) and physical books (photograph the cover). Maps content to curriculum nodes. While reading, I can tap and speak — the app knows which chapter I’m in. I can photograph a page (finds page number, OCRs text). These captures are lightweight enough not to interrupt reading. One knowledge item per curriculum node, not per chapter — a second book on the same topic adds a source, not a duplicate.
Go for a walk, talk about what you know. Transcription → quality gate → dedup → analysis against curriculum and book sources → captured facts, missed facts, interesting connections, wonderings. Wonderings get routed to microlearning automatically.
But the real value goes beyond grading. Those transcripts contain far richer data than a score — my actual words, the connections I made, the things I found interesting, my organizing framework for the topic. Recognizing this led to the knowledge profile system.
Every voice transcript is chunked into atomic pieces — individual captured facts, interesting connections, wonderings, the raw speech itself — each embedded as a vector and linked to curriculum nodes and entities. When I talk about the Fourth Crusade during a Constantinople elicitation, those chunks get linked not just to the Constantinople node but to every node and entity I mentioned: the Venetians, the Latin Empire, Enrico Dandolo.
This creates two retrieval strategies working in parallel. Relational: which chunks are directly linked to this node? Semantic: which chunks across all my transcripts are most similar to this topic, even if I never explicitly connected them? When the system generates a review card about Aristotle, it now finds my words about the Alexander-Aristotle relationship, my confusion about Nicomachus’s role, my interest in French neoclassical drama’s reception of the Poetics. The LLM can reference my own connections, address my specific misconceptions, and build on what I found interesting rather than generating generic questions.
For domains with enough elicitation, the system synthesizes a domain knowledge portrait — a 300–500 word summary of my mental model, strong areas, gaps, interests, and organizing framework. This portrait is injected into every LLM prompt for that domain, so the system’s understanding of me persists across sessions and deepens over time. It’s not a score. It’s a rich profile that gets more useful with every voice interaction.
The knowledge profile also fixes a practical problem: candidate selection. Before, if I discussed the Fourth Crusade extensively during Constantinople recall, the system would still offer “The Fourth Crusade” as a fresh topic because it only checked which nodes had direct elicitations. Now it checks cross-node coverage through chunk links, so topics I’ve already covered tangentially are deprioritized rather than repeated.
Generated from books, voice, and microlearning. Enriched with temporal hooks, multi-source context, cross-curriculum connections. Factual questions generated deterministically from structured data; LLM only for contextual answers and analytical questions. After grading, “Also want to know” chips expand sideways.
When a wondering or follow-up triggers research: a structured card with primary sources, material evidence, places you can visit, and sideways follow-up questions. Each card generates quiz questions on expanding intervals. A wondering spoken during a drive on Tuesday becomes a researched card by Wednesday, quiz questions by Thursday, actively maintained by the following week.
I don’t want an iPad app for reading PDFs with highlighting tools. During my PhD, I built elaborate pipelines — downloading papers, highlighting, extracting highlights, processing them. That worked as a full-time student. It doesn’t work for an adult with a family.
I barely open my laptop outside work. I try not to look at my phone much when I’m home with my family. I treasure the hours of deep reading — by the fireplace early on a Saturday morning, or half an hour in bed. Those are sacred. The system should never make me feel like I should be annotating or feel guilty for not engaging with the app.
The same constraint shaped Alif. The way I learn Arabic might not be ideal for a full-time student, but the question isn’t what’s theoretically optimal — it’s what will actually work in my life. 5% of my time with the app, 95% with books, podcasts, museum visits, historical movies. But that 5% should significantly improve retention, the ability to build on what I read, and the enjoyment of future reading.
The idea is simple: review information at expanding intervals — 1 day, 3 days, a week, a month. Each successful recall lengthens the interval. It started with physical flashcards and Leitner boxes in the 1970s, then went digital with SuperMemo and Anki. Medical students use it to memorize anatomy; language learners use it for vocabulary. It works spectacularly for discrete facts.
Andy Matuschak and Michael Nielsen pushed into more ambitious territory. Their key insight: learning has two stages — encoding (understanding something for the first time) and maintenance (keeping it accessible). Traditional SRS focuses on maintenance. With Quantum Country and the broader Orbit project, they embedded review prompts directly into prose, integrating both stages. Under 95 minutes of review produced 54 days of average retention. Users at technical seminars could follow 40–45 minutes of material that would otherwise have been impenetrable.
In the broader personal knowledge management space, others have been experimenting too: SuperMemo’s incremental reading (embedding cards within source context since 2000), Ryan Murphy’s fractal understanding, and various experiments with spaced attention — resurfacing ideas at intervals, not for drilling but for connection-making.
But Matuschak’s most important contribution was a pivot: “What seems like a problem of forgetting is sometimes a problem of reading comprehension — never having understood in the first place.” For complex reading, traditional flashcards struggle with arguments and causal chains. Even in his Great Books course: “I don’t know how to write useful prompts. It’s not details I care about, but the habits and mindsets I’m developing.”
I never managed to use flashcards productively — not for languages (until Alif), not for history. They felt disconnected from reading, the process felt like homework, and what I wanted to remember didn’t fit on a card.
What changed: integration. Review cards generated from books I’m reading, enriched with multi-source context and temporal hooks. Voice elicitation as free recall during walks. Microlearning that follows my curiosity. The expanding intervals are preserved — rock-solid. But the isolated flashcard is replaced by a curriculum node enriched by multiple sources, assessed through free recall, connected to what I already know.
An earlier project, Hamarquizen, was the direct precursor. I built it for an 11-year-old preparing for a local history quiz competition, inspired by Orbit. The challenge: help him retain a lot of factual knowledge about Hamar’s history in a few days. I created a curriculum of the 20 most important people and events, wrote micro-stories for each, and integrated quizzes that fed into a scheduling bucket — the PRIME → READ → TEST flow. Rich memory hooks tied facts together: Hanna Winsnes was born in 1789, the same year as the French Revolution. It worked — he learned a huge amount of facts in a few days and found it fun. That pattern — author-crafted prompts in narrative context, temporal hooks, bounded curriculum — carried directly into Petrarca’s microlearning design.
I’ve just started getting to where Petrarca feels genuinely useful day to day. It will need iteration and measurement over much longer — months, not weeks. Progress is harder to measure than Alif, where I can see 1,300 lemmas growing at a clear rate. History knowledge doesn’t have an equivalent metric. Some approaches I plan to test:
Voice knowledge sweeps. Monthly: “Tell me the history of Sicily, roughly chronological.” Same prompt repeated = comparable data. I can measure coverage, connectivity, accuracy, organization. Goldsmith et al. found that similarity between a student’s network and the expert’s correlates with exam performance at r = .74. The curriculum graph is the expert network; voice transcripts reveal mine.
Review card trends. Accuracy and interval growth by knowledge type. Are dates stabilizing? Are connections strengthening?
The knowledge portrait. The domain portrait described above is the most promising approach. Because it’s generated from my actual voice transcripts, it captures not just what I know but how I organize it — whether I think chronologically or thematically, which connections I naturally make, what I find interesting versus what I struggle with. Regenerating it after new elicitations creates a longitudinal record: I can diff my April portrait against my July portrait and see what changed. Did I fill the gap around the Investiture Controversy? Did my framework shift from biographical to more structural? The portrait also feeds back into the system — it guides which questions get generated, which topics get prioritized, and what the LLM assumes I already know.
Whether any of this produces a trustworthy signal of deep learning is an open question. The early signs are encouraging — review cards generated with learner context feel noticeably more relevant than generic ones — but I need months of data before drawing conclusions.
Single-user Arabic learning app. Sentence-first, FSRS scheduling. ~1,300 words in 8 weeks. The design principle: the best way to find something universal might be to make it perfectly specific first.
10 books, 47 years, 10,312 claims, 656 cross-book arguments. Hirsch’s thesis — knowledge is the foundation of literacy — directly informs Petrarca.
~128 Norwegian Green Party programs: 7,600 proposals extracted, clustered. Same extract-embed-compare architecture. Technical details.
Norwegian newspapers from 1814, day by day. History through primary source immersion.
Frontend: Expo SDK 54 (React Native), 2-tab + drawer. Backend: Hetzner VM, Python, 4-hour cron pipeline. Data: SQLite. LLM: Claude (Opus for curricula, Sonnet for routine), Gemini (interactive). Embeddings: Nomic-embed-text-v1.5 (768d, AUROC 0.930). Scheduling: FSRS-inspired, stability multipliers (knew 2.5×, partly 1.5×, missed reset), engagement-based initial stability (9d skimmed to 120d annotated). Voice: Soniox transcription, perceptual fingerprint dedup, <15 word quality gate. Named after Francesco Petrarca (1304–1374).
Nielsen, M. (2018). “Augmenting Long-Term Memory.” Memory as infrastructure for expertise.
Matuschak, A. (2019). “Why Books Don’t Work.” Transmissionism and metacognition.
Matuschak, A. & Nielsen, M. “How can we develop transformative tools for thought?”
Matuschak, A. Orbit. Open-source spaced repetition platform.
Wineburg, S. (1998). “Reading Abraham Lincoln.” Cognitive Science, 22(3).
Wineburg, S. (2001). Historical Thinking and Other Unnatural Acts.
Hirsch, E.D. (2016). Why Knowledge Matters. The France case.
Hirsch, E.D. (2024). The Ratchet Effect. Cumulative cultural evolution.
Sweller, J. (1988). “Cognitive Load During Problem Solving.” Cognitive Science, 12(2).
Miller, G. (1956). “The Magical Number Seven.” Psychological Review, 63(2).
Brainerd, C.J. & Reyna, V.F. (2005). The Science of False Memory. Fuzzy-Trace Theory.
Karpicke, J.D. & Blunt, J.R. (2011). “Retrieval Practice Produces More Learning.” Science, 331.
Goldsmith, T.E. et al. (1991). “Assessing Structural Knowledge.” J. Educational Psychology.
Recht, D.E. & Leslie, L. (1988). “Effect of Prior Knowledge.” Knowledge > reading skill.
Schneider, W. et al. (1989). “Domain-Specific Knowledge.” Low-IQ experts > high-IQ novices.
Rawson & Dunlosky (2022). “Successive Relearning.” J. Experimental Psychology: General.