Polymath

Polymath

Epiphany

I want to read everything. I want to capture everything that I can that I can read. There are a lot of good ideas floating around some original, some enhanced some stolen and some cloned but they are good and I want to organize this information into proper buckets and find smart and deterministic way to retrieve them and create relations and neural networks between these topics of data points. I want to build a neural mind palace of information which stores this supermemory of knowledge which is contextualized for me and tailored for me.

I want to think about ways to store this kind of information and how storing and perhaps later using (in the form of flashcards, quizzes or written summaries analyzed later) to get a report of where my own knowledge and understanding of topics stand and how I am performing and how I can be the best knowledge worker is what I want to achieve.

The thing is, sometimes there is not structure to the knowledge you have. You dont know how much you are aware of and you also dont know how much you need to be aware of. If only I had a progress indicator showing how much I am prepared for a meeting or for a coding interview. Not only that but it is also difficult to define the data and knowledge you have. Its difficult to measure it, it is difficult to put a shape on it. It is difficult to retrieve the exact kind of information required to deterministically or close to deterministically finish a task. If only you could visualize a task at hand, draw out knowledge and information to fill those.

Compartmentalization of the brain and tasks is very important. I want to build a deep focus UI/UX experience that lets a user stay hyperfocused on the screens unless and until they are contextually doing the task and have a screen reader drowsiness detection system that detects slacking off. That will be part of Dao.

I want a dual factor application where I used Dao and Alf side by side and I want Alfred to be my master backend for knowledge management and Dao also obviously has a lightweight nodejs backend.

I want Dao to be the work automation platform and Alf to be a knowledge automation platform and work and knowledge work together to give one a complete sense of productivity and fulfilment. Life on the screen is stressful, giving users a nice interface to constantly work with would be a good idea.

Ideas for features

Spaced repetition

Active Recall - After a section, close the book and write or say what you remember; testing yourself strengthens memory far more than re-reading.

Elaboration - Explain ideas in your own words and connect them to what you already know; asking “why” and “how” creates richer neural links.

Dual Coding - Combine words with visuals (diagrams, timelines, sketches) so the same idea is stored in multiple formats, improving recall.

Interleaving - Mix related topics during study sessions instead of blocking one topic for hours; this improves discrimination and transfer.

Feynman Technique - Teach the concept aloud to an imagined novice; gaps in your explanation reveal what to study next.

Understand the Purpose - Determine the author's intent and scope.

Checking Vocabulary - A specialized dictionary. UI deep research with small Ipad like Dictionary. This can be complemented by keeping a vocabulary journal.

Summarize key points - Write and summarize the key points learned.

_Zettelkasten like view and infact the ability to view information in different views. Graph based views.  Zettelkasten. Diagrams. Different ways of representing data._ 

This is an applied Epistemology project

All of this is basically epistemology — the philosophy of knowledge itself — but turned into software. The problems I'm describing above are the same ones philosophers have been wrestling with forever:

• "I don't know how much I know" → thats metacognition. The readiness indicator is just know thyself but quantified.
• "Its difficult to measure knowledge" → Bloom's Taxonomy says remembering something is a lower form of knowing than being able to apply or create with it. Alfred's dashboard cant just track "have you seen this" — it needs to track "can you use this."
• "Deterministic retrieval" → philosophers call this epistemic access. Can you get to the right knowledge at the right time? A knowledge graph with semantic search is an engineering answer to that.
• Spaced repetition, active recall, Feynman Technique → these are all ways of moving from just believing something to actually knowing it — strengthening the justification leg of the knowledge triangle.
• Empiricism (Locke, Hume) - Knowledge comes from experience. Relevant to an idea of learning by doing.
• Constructivism (Piaget, Vygotsky) - Knowledge is actively built by the learner, not passively received. This is the theoretical basis for why active recall and elaboration work.
• Connectionism - Knowledge is stored as patterns of activation across networks of nodes. This literally describes both how your brain works and how your knowledge graph in Alfred would work.
• Personal Knowledge (Michael Polanyi) - We know more than we can tell. Tacit knowledge resists codification. This is the hardest problem for Alfred - capturing what you know but cannot articulate.
• The Daily Collision. Every day, Alfred takes two notes from different domains and presents them side by side. Not randomly — semantically. It uses embedding similarity across domain boundaries. Your note on "LangGraph's Send() pattern for dynamic fan-out" and your note on "how ant colonies distribute foraging tasks" — both are about parallel task distribution, but you filed them under AI Engineering and Philosophy respectively. Alfred surfaces the resonance. You either dismiss it (not useful) or link it (creates a new edge in your graph). Over time, Alfred learns which cross-domain collisions you find valuable.
• The Contradiction Detector. Alfred should surface tensions. "In your note from January, you wrote that microservices reduce complexity. In your note from March, you wrote that distributed systems increase operational overhead. These seem to be in tension. Can you reconcile them?" This is where real understanding lives — in the resolution of contradictions.

The Core Interaction Loop

The magic isn't in any single feature. It's in the loop:

Capture → you ingest something (article, book chapter, conversation)

Atomize → Alfred helps break it into zettels (or you do it manually)

Connect → you or Alfred link zettels to each other and to domains

Test → spaced repetition, active recall, Feynman prompts challenge your understanding

Measure → your dashboard updates with coverage, depth, freshness

Apply → Dao sends a task, Alfred surfaces relevant knowledge and gaps

Strengthen → knowledge used in real tasks gets reinforced; gaps get flagged for study

Stuff I want to get better at - curriculum

• System Design
  • Designing Data Intensive Application
  • Alex Xu All Lessons
  • Systems thinking
  • Backend Engineering
• AI Engineering
  • Langgraph docs
  • Anthropic / OpenAI / Xai / Ramp / Stripe Blogs
  • Claude Code
• Finance
• Startups
• Investments
• Writing and Literature
• Politics
  • Geopolitics
  • Current affairs
  • Global Policy
  • AI and Tech
  • HealthTech
• Movies and Pop-culture
  • Where to find
• Philosophy
  • What to read

What I’d Like Alfred to be

Essential what we are describing maps to a well-studied framework in knowledge management called the DIKW Pyramid (Data → Information → Knowledge → Wisdom). Most "second brain" tools stop at Information (storing and retrieving). The leap to Knowledge and Wisdom requires active synthesis — connecting dots, spaced repetition, progressive summarization, and contextual recall. Alfred has the raw ingredients today here: but they're currently siloed. I want Alfred to

1. Capture all data in a unified Mailbox
2. I want to see a knowledge graph / concept graph / zettels / different views and relations of my knowledge
3. ofcourse “chat” with my knowledge, rag, chatbot, vector, agents, ai, agi, nvidia, claude-code
4. A dashboard to measure of much i know of what is in alfred
5. creativity and the concept of god is all the knowledge in the universe

How can I embed a sense of measuring my own understanding of this ultimate knowledge base?

Good question for claude code.

Market Research

What actually goes wrong

We did the homework. Most tools in this space are built on intuitions about human behaviour that turn out to be wrong. So this section is the evidence base for every design choice below: each item names what the research found, and what Polymath does differently. Where the research argues against us, we write that down too.

2.1 People save far more than they retrieve

The finding. Abrams, Baecker and Chignell surveyed 322 web users and analysed 50 bookmark files. The median bookmark had not been visited in about 100 days. Roughly half had been visited in the previous three months; 67% within six months; 96% within a year (CHI '98).

The anecdote. Someone reads a useful article during a project, saves it "for later," never opens it again, and six months later finds the same article by searching the web. The saving cost two seconds and bought nothing.

2.2 Re-finding mostly does not use the archive

The finding. The Keeping Found Things Found project found participants re-found 95% or more of pages after three to six months - and two-thirds of successful re-findings used no explicit keeping behaviour at all. Search engines and URL autocomplete did the work (Jones et al., 2004).

What this means. Elaborate capture does not pay off, and the reason is not that the capture is bad. The archive was never the mechanism for re-finding.

Therefore. Polymath does not compete with a search engine on recall. It stores what a search engine cannot reconstruct: what you concluded, what it connected to, and what you decided to do about it. A captured source is the evidence behind a conclusion, not a bookmark.

2.3 Systems die of maintenance cost

The finding. Jones, Bruce, Klasnja and Jones ran a six-month interview study and identified five abandonment factors: visibility, integration, co-adoption, scalability, and return on investment - the system must pay back faster than it costs to maintain (2008).

The anecdote. A person builds a careful tagging taxonomy over a weekend. Three weeks later one misfiled note makes the taxonomy inconsistent, fixing it is expensive, and the system loses credibility. They stop tagging. The archive becomes a pile.

Therefore. Capture requires zero decisions. Classification is proposed with a reason and accepted in one action. Linking is optional and rewarded, never required. A library with ten links is not a failure.

2.4 Interruption costs stress, not only time

The finding. Mark, Gonzalez and Harris observed 24 information workers: an average of 11 minutes 4 seconds in a working sphere before switching. 57% of spheres were interrupted; 77% were resumed the same day after an average of 2.3 intervening activities (CHI '05).

The finding. A 2008 follow-up (N=48) found interrupted participants finished faster with no quality difference, but reported significantly more stress, frustration, time pressure, and effort (Mark, Gudith & Klocke).

A misquotation we refuse. The widely repeated "23 minutes 15 seconds to regain focus" is not in Mark's papers. It traces to a 2006 Gallup interview describing same-day return to the original task, not time to regain depth. The 2008 paper's abstract says the opposite of how the number is used (forensic account). Polymath does not use this number.

The finding. Leroy (2009) showed "attention residue": unfinished tasks keep consuming attention after a switch (OBHDP).

Therefore. The focus feature is framed honestly. Polymath does not promise to eliminate switching or make you faster. It promises to make returning cheap: one task, its intention, and the last thing you did, restored in one action. The benefit is lower stress and fewer lost threads.

2.5 Spaced repetition works, but narrowly

The finding. Dunlosky et al. reviewed ten study techniques. Only practice testing and distributed practice received "high utility." Highlighting, rereading and summarisation were rated low utility (Psych Sci in the Public Interest).

The finding. Cepeda et al.'s meta-analysis covered 839 assessments across 317 experiments: the optimal spacing gap scales with retention period (Psych Bulletin). A later study found the gap should be 20-40% of a one-week delay but only 5-10% of a one-year delay (Psych Science).

The finding. Karpicke and Roediger showed repeated testing after apparent mastery produced large delayed-recall gains, while repeated study produced none (Science).

Therefore. Cards are generated from notes and claims, scheduled by expanding intervals, and tested by retrieval. The product claims improved recall of facts. It does not claim improved judgement or decision quality, because the evidence does not support that.

2.6 Rewards help quantity and can harm motivation

The finding. Deci, Koestner and Ryan's meta-analysis of 128 studies found engagement-, completion- and performance-contingent rewards undermined free-choice intrinsic motivation (d = -0.40, -0.36, -0.28). Positive feedback enhanced it (d = 0.33) (Psych Bulletin).

The finding. Mekler et al. found points, levels and leaderboards acted as extrinsic incentives, effective for the quantity of performance but not for intrinsic needs (Computers in Human Behavior).

The finding. Moldon, Strohmaier and Wachs studied GitHub's silent removal of public streak counters: long streaks were abandoned, weekend activity fell, and some developers quit in-progress multi-day quests outright (ICSE). The cleanest natural experiment on streaks.

The finding. Renfree et al. found streaks and reminders work by creating dependency: when users abandon the app they also disengage from the behaviour it scaffolded (CHI EA).

Therefore. Polymath rewards completion, never continuity. There is no streak. Nothing decays, nothing is lost, and a week away costs nothing. Stones are never revoked. Shards are claimed, never auto-converted.

Correction to the current code. src/domains/gamification defines a streak_day reward kind. It contradicts P-11 and the evidence. It should be removed, not used.

2.7 Document-heavy work is bottlenecked on human reading

The finding. RAND studied 57 large e-discovery productions: median total spend $1.8M. Review consumed a median 73% of cost, at roughly $18,000 per gigabyte. Automated deduplication and filtering reduced volume by only 34% (RAND, Where the Money Goes).

The finding. Panko's work on spreadsheets found errors in a few percent of all cells, meaning large models almost certainly contain at least one wrong bottom line, and developers are systematically overconfident about their own sheets (1998).

Therefore. The professional-services persona is real and the pain is documented - but the value Polymath offers is skimming and citation, not autonomous judgement. Every extracted claim carries a citation to a span so a professional can check it.

A number not to use. McKinsey's finding that 28% of the workweek goes to email and about 20% to searching internal information is real. The popular "1.8 hours a day searching" restatement is derived and frequently misattributed. Do not cite it as measured.

2.8 The honest counter-case

The strongest arguments that a product like this should not exist. Recorded so the product answers them rather than ignores them.

1. Re-finding does not need an archive. (§ 2.2) If search already works, capture must justify itself by storing conclusions, not sources.
2. Folders beat the theory. Bergman and Whittaker find hierarchical folders outperform their critics' claims because they encode where as well as what; search and tags underperform on personal collections (MIT Press).
3. Systems die of maintenance cost. (§ 2.3)
4. A perfect archive may make you less knowledgeable. Sparrow, Liu and Wegner found that expecting future access reduces recall of the information itself and increases recall of where to find it (Science 2011).
5. Learning evidence does not transfer. Everything in § 2.5 concerns factual recall. Nothing shows better organisation improves synthesis or judgement.

Polymath's answer. The product is judged on decisions and finished work, not collection size. If a feature grows the archive without shortening the path to a decision, it is not built. That is the test this document applies to itself.

3. Who it is for

We build for a person, not a market. Here is that person, drawn honestly - including who we are not for.

3.1 The ideal user

Works largely alone. Produces written or analysed output. Reads more than they write. Keeps notes somewhere they do not revisit. Uses LLMs daily. Carries recurring obligations with deadlines.

3.2 Personas

Persona A - The generalist operator

A software engineer who researches constantly, writes technical long-form, prepares for interviews, and manages job applications.

User story. "I read four articles while fixing one bug. Two weeks later I write a post about that bug and cannot remember which article said what." Polymath keeps the source next to the note, and the claim carries the citation.

User story. "I send eleven emails a day and half are variations of the same three messages." Polymath drafts from prior context; the person approves before anything sends.

Persona B - The professional-services principal

An accountant or lawyer whose work is intake, classification, and judgement over documents, with consequences for being wrong.

User story. "I retype transactions from a bank statement into a ledger and classify each one. I cannot afford to misread a figure." Polymath proposes a classification against the source with the figure visible, and never commits without showing what it read.

User story. "A client sends a 60-page contract and asks whether a clause is standard." Polymath cites the span it is answering from, and states when it does not know.

Persona C - The researcher-writer

Produces papers, briefs, or essays from many sources, where an unsupported claim is a real failure.

User story. "I have 40 sources and need to know which support this argument and which contradict it." Every claim carries its source and confidence; uncited claims are visibly marked.

3.3 Who it is not for

• Teams coordinating on shared documents. v1 has no real-time collaboration.
• People who want an archive to browse. The product optimises for finishing.
• Anyone needing regulatory-grade assurance from automation alone. Polymath proposes; a human disposes.

Competition

Every tool in this space picks 2-3 features and ignores the rest. Nobody does it all.

Obsidian — The power user favorite. Local-first markdown notes, bidirectional links, knowledge graph visualization, 1000+ community plugins. Spaced repetition only via plugin, no native AI, no measurement. Graph view looks cool but most people say its more distracting than productive.

RemNote — Closest to this vision. Notes + flashcards + spaced repetition natively integrated. Turn any note into a flashcard, SM-2 algorithm, daily review queue. But no knowledge readiness dashboards, no "how much do I know" measurement, limited graph views.

Heptabase — Visual-first PKM built around infinite canvas whiteboards. Card-based notes on spatial canvases, nested whiteboards, AI research assistant with citations. No spaced repetition, no flashcards, no measurement. YC backed.

Mem — AI-native note-taking. Auto-organizes and surfaces relevant notes, chat with your notes (RAG), voice capture, web clipper, meeting transcription. No spaced repetition, no knowledge graph, no measurement dashboards, no Zettelkasten views. SOC 2 Type II.

Tana — Outliner with "supertags" that add structured fields/schemas to notes. AI commands for generation. No graph view, no canvas, no spaced repetition, no measurement.

Logseq — Open-source outliner with graph view and native flashcards/spaced repetition. Whiteboards, block-level references. Flashcards are basic compared to RemNote, no knowledge dashboards, limited AI.

Read AI — Builds a personal knowledge graph from meetings, emails, messages, docs. Cross-channel intelligence. Not really a PKM tool though — more of a workplace intelligence layer.

Where Alfred wins

The biggest gap in the market is measurement. Every tool helps you capture and organize. RemNote and Logseq add retention via spaced repetition. But nobody answers "how much do I know?" — thats the unique angle.

• Knowledge readiness dashboard → nobody has this
• Bloom's taxonomy progression tracking → nobody has this
• Curriculum coverage scoring → nobody has this
• Active recall + Feynman testing built in → nobody does this well
• Multiple views (Zettelkasten, graph, cards, diagrams) + spaced repetition + AI chat + unified inbox → nobody combines all of these

The closest analogy is what Duolingo does for language learning (progress bars, streaks, skill trees) — but generalized across any knowledge domain.

Where we are today - May 3rd 2026

Today we have renamed Alfred to Polymath. Polymath is an incubation center for ideas. For the ever evolving thoughts that you capture when you do knowledge work. Inspiration can be sought from anywhere, inspiration can be gotten from anywhere. You want to shape this inspiration in a proper a way when it comes to you and store it. A idea could be like a diamond in the rough, but this diamond may only have value until given a polish. This is what polymath’s intention is to be.

This opens up so many avenues to achieve what can be achieved from Polymath, but we will for the time being stick to an idea incubation center and see what all it makes possible in one self - the changes in the minds that can take place with the behavior change introduced by this product.

Some brainstorming session QnA with an AI agent

---

Q1 — Demand Reality

What's the strongest evidence you have that someone besides you actually wants Alfred — not "is interested," not "would try a waitlist," but would be genuinely upset if it disappeared tomorrow?

Concretely, I'm looking for one of these, in order of strength:

1. Payment: someone paid. Even .
2. Workflow dependency: someone restructured their daily reading/thinking around it, and would scramble if it vanished.
3. Panic: someone got angry or anxious when it broke (an outage, a lost note, a bad migration).
4. Sustained unpaid use: a non-you human opened Alfred more than 5 days a week for 3+ weeks without you prompting them.

If the honest answer is "none of the above, I'm the only user" — say so. That's the most useful answer possible, and it reframes the entire session. I would rather hear "zero external users, I haven't shown it to anyone" than a fuzzy "a few friends have signed up."

Im the user currently.

I can see content creators using this application. Polymath content creators or people interested in expanding their horizon to learn more things and people wanting to create with those things. For example there are many people I know who are interested in reading about different things - technology, art, movies, inspiration and they might want to write about it or blog about it or scriptize it somehow. I have often found these curious people in need of organizing their thoughts and where they get their inspiration from.

I can see students clearly using this application as a professional - notetaker in class. Something that lets students learn and think for themselves (Today Polymath might not be an interface which makes learning easier) A classroom is a place where ideas are easy to be brought - challenged by a peer, reviewed, analyzed, discussed, pros-and cons and related ideas better or worse ideas. So a classroom is a place where it is easy to have such ideas have such treatment either from other students or from libraries or teachers. So we need to build such setting (for example [a small example] - the sitting session in Zettelkasten service needs to be interactive and stream relatede content and always guide user as to what they are learning and what is the related content they read always in real time.

And many more. You tell me.

Q-Builder 1: The Personal Moment Test

Tell me about the most recent time — within the last two weeks — that you opened Alfred and it failed you in some way. Not a bug. Something more subtle: a moment where you wanted Alfred to do something, you reached for it, and the tool either didn't have the capability, had it but buried it, or had it but did it badly enough that you fell back to something else (Notion, Chrome tabs, a text file, your memory).

This question is the n=1 equivalent of Q1 — it's asking for evidence, but evidence about you. And it's the most important question in this session,

because the gap between "Alfred as it runs" and "Alfred as you actually use it" is the real design space.

Give me one concrete story. Shape of answer I'm looking for:

▎ "Last Thursday I was reading [thing] and wanted to [action]. I went to [Alfred surface] and [what happened/didn't happen]. I ended up [fallback].

▎ What I actually wanted was [capability]."

If you cannot think of a moment — if Alfred is serving you perfectly today and you only reach for it and it always delivers — that's also a real

answer. It means the gaps are elsewhere (in ambition, in surfacing, in connection) and we'll adjust the next questions accordingly.

One story. Recent. Specific. Your turn.

Well the Universe view is abysmal. That is something that I wanted to enjoy with an almost gamelike speed experience where I drift through connected knowledge zettels swiftly and ccan have AI chat and explore related zettels and topics well. The reactflow graph view also does not impress me too much. I wanted this view to be the core thesis on how I can view my knowledge and its connections in different ways. I like the basic facade of the universe view but it is missing that magic touch and snappiness. It also can do with an agentic chat within itself idk. I want dopamine boost while using this application.

Dashboard

May 24th Polymath Additional Features

I want to create a new way of keeping focused. I want to find a way of abstracting away deep focus or deep work. The cal newport way of just keeping on focusing on something until it is done. So I want to ask myself then that is there a way to abstract away what we have in Polymath or could have in alfred and with the neural architecture and design’s, create for a Polymath a way to keep focused on work. For example here is how it looks like for me, myself -

I do all my work on Computer Screens. I measure my work on browser tabs.

Check the page above which has 9 open browser tabs. I expect some sort of mechanims to be aware of tasks on my interaction or on autmation (I want the perfect balance) and I want us to create some todo lists automatically. That would be nice. It would atleast make me control my chaos and then depending on the information on the page I can decide what level of information I want to consume - I have the inbox capture which can summarize and enrich - I can just asynchronously process those pages or tasks and enrich and organize them and read them and after reading them I mark them off as done. I want to call this the todo system that automatically creates these todos in the today page for me. I want today to be an entry point for a master command center audit trail of my work throughout the days. I also want it to be a motivational and frontend for notification system updates in there wherin i am notified of pending tasks, what needs to be done, what got done (with strikeoffs),

The Notification and Nudging Aspect

I want a smart notification sidebar system (where tasks are displayed currently I think that is a good sidebar for notification client) but I want the notification system to be aware of the system and what needs to be notified to me. Choose the best and wisest architecture here in the backend that best suits the needs.

The Gamification Aspect

I want then to extend this feature into a Gamification aspect wherein I am using the stone design gem design award principle when tasks are finished hence I want to motivate myself to do knowledge work well. I want a proper reward system setup that rewards gemstone images into my vault on task completion. Create a robust and smart system that acts up on the nudges of the notification system and the todo system.

The Newsletter Aspect

References

The source provides a comprehensive overview of the Zettelkasten, a German term for a card file system used for personal knowledge management and scholarly note-taking. This method relies on creating small, individual slips of information (Zetteln) that are linked together through subject headings, tags, and cross-references, enabling deep organization and idea retrieval. The article traces the technique's rich history, from early adopters like Conrad Gessner and Carl Linnaeus, to its modern use by highly prolific thinkers such as sociologist Niklas Luhmann. Ultimately, the text establishes the Zettelkasten not just as an analog tool, but as a foundational concept that influenced the development of early hypertext systems and modern digital wikis, highlighting its power for enhancing creativity and research output.

More References and Ideas Dump

What is Built So far

More Ideas Gleaned

The Meaning of Self-Knowledge

In the concluding sections, Jung argues that the salvation of the world does not lie in mass action, political programs, or external social reforms, but in the psychological transformation of the single individual. He stresses that individuals who do the difficult work of accessing their unconscious and achieving self-knowledge naturally exert an "unconscious prestige" (or mana) that positively influences those around them

This is basically what I want Polymath to be a “mana” a hindi word which was there right in front of my eyes. Man ko sab pata hai - the mind knows everything.

Polymath is meant for selective knowledge exposure for focused task execution too