The idea

The space of words that grow and relate to one another.

wordspace is the space where your words grow and relate to one another. Not an archive: a fabric that forms by itself, as you write.

wordspace works on two materials. What you write from today on becomes a wiki: that is the first part of this page. What you have already written over the years — the backlog, scattered across your folders — is found again in Themes: that is the second. Go to Themes ↓

Why tidying shouldn't be your job

Taking notes is easy. Finding them again, linking them, noticing that two notes contradict each other — that is the work no one has time for. wordspace does it for you, as you write. You think in notes; the intelligence, entirely on your Mac, re-reads them and organizes them by topic, people and ideas.

None of the model's words enter your wiki: only what you write does. The model observes; it does not dictate.

The same holds for what arrives from outside. A voice note caught in the street, a prompt the app suggests: they stay beside your note, readable, but they don't become your thinking. What enters the wiki is what you write after passing through them.

How it works

Every time you save, a local engine re-reads the new or changed notes and updates the map.

The loop closes: when you answer a contradiction or capture a prompt, what you write comes back in as a note. The system doesn't hand you back what you know — it makes you write what you didn't know you thought.

Three numbers

Contradictions are not hunted by the model: they are hunted by the code, and the model steps in only where understanding a language is required. It is the design choice that makes everything else possible, and it is measured.

Zero false alarms in 18,000 casesThe deterministic comparison of facts — amounts, dates, quantities — produced 7,333 flags in six languages without a single mistake. It doesn't depend on which model you installed: it's code.2.73 GB in memoryLess than a browser with a dozen tabs. Everything fits inside: extraction, rules, judgement, summaries, search.144 chances to invent a number, none takenHalf the test questions had no answer in the notes. Neither model answered anyway.

The full study also says where the system doesn't hold: the measurements, in full. →

Benefits

Intelligence on the device

No servers of ours, no account, no tracking. The model runs on the Mac and your writing never reaches us: between iPhone and Mac it travels in your iCloud.

It takes the tidying, not the thinking

Topics, people and links emerge from the notes. You don't build the structure — you find it already made. The time saved goes back where it's needed.

Thinking in dialogue

The wiki shows you the contradictions between things you wrote at different times, and asks you positioning questions. Not answers: questions.

It matures over time

The more you write, the denser the web. Knowledge doesn't pile up — it connects.

It stays yours

Your texts are Markdown, an open, non-proprietary format: what you write is what gets saved, with no hidden fields and no closed formats. The wiki pages are readable files in the app's folder, and the app keeps a history of your notes' earlier versions.

Every claim carries its source

Quotations are checked character by character against your notes before they appear. The ones that don't hold are stopped.

What it isn't

wordspace is not an assistant that works in your place. It will hand you contradictions to resolve and uncomfortable questions, and neither resolves itself.

Nor is it an archive that merely stores. It notices when you change your mind, when an external source is silent on exactly what matters to you, when a concept recurs often enough to deserve its own page.

It's a tool to think better, not to hoard more. The sign that it's working isn't that it seems clever: it's when it shows you a friction that bothers you.

A friction, though, only bothers you if it is real. An invented contradiction is as irritating as a real one, and after the second or third you stop looking at any of them. That is why, in our work, precision matters more than how much you find: a conflict you miss costs an opportunity, a conflict you invent costs the tool.

The name says it: wordspace is the space of words — where words grow and relate to one another.

The archive you already have: Themes

From archive to themes: finding your work among everything else.

A twenty-year archive can’t be tidied up: you discover what it contains. This tab looks at the folders you choose, takes what you wrote yourself, and groups it into themes. You choose which ones to keep, what to keep inside each theme, and what to bring into wordspace.

The problem

Folders say where you put things, not what they’re about. One case sits in three places, a “Misc” folder holds ten, and most of the files weren’t written by you: they’re received deeds, forms, attachments, downloads.

And nobody decides on forty thousand rows. The job isn’t trimming a list: it’s finding the core of your own things inside the mass, and deciding on that.

How it works

No step moves, deletes or writes a file. Only what you bring in enters wordspace — a document into the Inbox or a note, the notes of the themes you import: on its own, the tab writes nothing among your notes.

Inside a theme

A big theme isn’t kept or discarded whole. Inside, it has three levels of decision, all by whole groups — none file by file.

Keep or archive — the whole theme, with a switch. Archiving doesn’t delete: the theme moves down to the “Archived” section and can be switched back on from there.
Subthemes — the groups that emerge on their own inside the theme. They’re shown only if they hold up: if they change from one sample to the next they’re noise, and none are shown.
Sources — internal — written by you and not duplicates of others — and external: deeds, forms, attachments, duplicates. They’re kept or left out in groups.

The file list below is always a portrait of what you keep: it drops on its own whatever the choices above leave out.

Two ways to look at a theme

As text — subthemes, sources, files — or as a cloud: every document as a dot, laid out by folder, with the theme’s documents lit up. In the cloud you see at once whether a theme sits in a single folder or is scattered everywhere; hovering over another theme in the list lights up its documents. You choose under “View”, and it applies to every theme.

What’s measured

Every threshold in the tab was calibrated on a real archive or on archives built on purpose, where the right answer is known. A few examples, including the one that only partly works.

From 61% to 74% of your documents inside a themeThe groups were made of short forms, and the long documents — the ones really written — were left out: they were four times longer than those inside. Length normalization brought them back in.

The largest row from 38% to 8% of the archiveA theme holding a fifth of everything you wrote is shown broken into its subthemes: a row containing a third of a lifetime’s work isn’t a choice.

A theme’s core from 93% to 40%Inside a theme the file format is almost constant, so on its own it doesn’t separate your documents from forms. What separates them is how much a document resembles its neighbors: a form has twenty twins, something written only once has none.

“Written by you”: from 69% to 84% correct judgmentsThe file format alone is often wrong: of the hundred documents annotated by hand, fewer than half of those it calls yours actually are. “Go deeper”, on a theme’s page, has a light local model read the beginning of each document, and together with the format it reaches 84%. It remains the weakest point, and that should be said.

The first analysis from 4 minutes to 45 secondsOn this archive — forty thousand files, seven thousand read — reading stays sequential, because the Word readers can’t handle concurrency, but text preparation uses all the cores. After that, only what has changed is read again.

What it doesn’t do

It doesn’t read the folders you exclude — and the tab always says which ones they are, even when you exclude none.

It doesn’t touch the files: nothing is moved, renamed or deleted.

It doesn’t decide for you. It proposes themes, cores and subthemes; keeping or archiving is your call.

It doesn’t cover everything. About a quarter of your documents resemble no theme: it’s the long tail of every personal archive, stuff put away and never looked at again. It stays where it is, and the tab tells you so instead of hiding it behind a nicer count.

The bridge between the two parts

The two parts stay separate: Themes don't write notes, and the wiki doesn't read the archive. You are the one who crosses the bridge. From a theme's page, “Import…” brings a document into the Inbox to wait, or into a note: as an external source, to consult but kept out of the wiki, or as a linked internal source, whose text goes into the wiki and follows the file.

And if a topic doesn't emerge on its own, “New theme…” opens one starting from a word that appears in your documents, together with the words that go with it.

References

The wiki

Vannevar Bush — As We May Think (1945). Imagines the Memex, a desk that stores books and notes and lets you bind them together into associative trails. The insight that matters isn't the storage: it's that the value lies not in the documents but in the links a person draws between them, and that those trails should be preserved like the material itself.
Niklas Luhmann — Communicating with Slip Boxes (1992). Describes the Zettelkasten with which he wrote dozens of books: short cards, each with one idea, linked to one another instead of filed by topic. The point is that the structure isn't designed in advance — it emerges from the links. And the archive, once grown large enough, becomes an interlocutor that returns juxtapositions you hadn't foreseen.
Sönke Ahrens — How to Take Smart Notes (2017). Brings Luhmann's method back to daily practice. The useful distinction is between fleeting notes, notes on what you read, and permanent notes written in your own words: only the last count, because rephrasing is already thinking. Taking notes isn't for remembering, it's for forcing yourself to understand.
Joon Sung Park et al. — Generative Agents: Interactive Simulacra of Human Behavior (2023). Introduces reflection: the agent doesn't just accumulate observations, but periodically re-reads them and synthesizes higher-level beliefs. It's the difference between a memory that grows and one that consolidates — and the recognition that without that step, volume smothers meaning.
Andrej Karpathy — llm-wiki (2026). Proposes compiling notes just once into a wiki of linked pages that the model maintains, instead of retrieving fragments at every question. The model reads, distills, updates and flags conflicts; knowledge accumulates instead of being rediscovered each time. The markdown stays readable so the person can check what the model wrote.

What each of these works receives in return from our measurements — and, on one point, where our measurements say the opposite — is at the end of the study. →

Themes

H. P. Luhn — The Automatic Creation of Literature Abstracts (1958). The words that matter are the ones that come up often in a document but not everywhere. It’s the starting idea behind all the word counting in the tab.
Karen Spärck Jones — A Statistical Interpretation of Term Specificity (1972). IDF: a word weighs less the more documents contain it. It’s the weight used for theme names and for the tree’s “words that come up often here”.
G. Salton, A. Wong, C. S. Yang — A Vector Space Model for Automatic Indexing (1975). Every document becomes a vector of weighted words, and two documents are as similar as their vectors. It’s how the tab decides what belongs together.
J. S. Justeson, S. M. Katz — Technical Terminology (1995). Multi-word expressions — “judicial mortgage”, “gross floor area” — say more than single words. They’re the minimal phrases at the head of theme names.
F. Murtagh — Hierarchical Clustering Algorithms (1983). The reciprocal nearest-neighbor chain: two documents join when each is the other’s nearest. It’s the algorithm that builds themes and subthemes.
R. Campello, D. Moulavi, J. Sander — Density-Based Clustering Based on Hierarchical Density Estimates (2013). Groups that stay stable across many heights of the tree are kept, not those from a single cut. It’s the criterion by which subthemes are chosen.
P. J. Rousseeuw — Silhouettes (1987). How much better a document fits in its own group than in the neighboring one. It’s used to choose how many themes to keep instead of fixing the number in advance.
A. Singhal, C. Buckley, M. Mitra — Pivoted Document Length Normalization (1996). Cosine similarity penalizes long documents, which spread their weight over many words. Correcting for it, the long documents that were left out of themes — the ones really written — came back in.
M. Grootendorst — BERTopic: class-based TF-IDF (2022). A group’s name is found by comparing its words with those of the other groups, not with the whole archive. It’s how themes get their names.
J. Carbonell, J. Goldstein — Maximal Marginal Relevance (1998). Choosing items that are relevant but different from those already chosen. It’s the principle by which a document with twenty twins counts less, in the core, than one written only once.
A. Strehl, J. Ghosh — Cluster Ensembles (2002). Normalized mutual information: how much two groupings say the same thing, from zero to one. It’s the measure used to compare resamplings.
A. Ben-Hur, A. Elisseeff, I. Guyon — A Stability Based Method for Discovering Structure in Clustered Data (2002). A grouping is real if it reappears the same on different samples of the same set. It’s the test a subtheme must pass to be shown.
J. J. Rocchio — Relevance Feedback in Information Retrieval (1971). A query grows with the words of the documents that answer it best. It is how “New theme…” suggests the words close to yours.
A.-L. Barabási — The Origin of Bursts and Heavy Tails in Human Dynamics (2005). Human activity comes in bursts, not at a steady pace. That’s why time tells a case — many documents in a few months — apart from a collection spread over years.
T. W. Malone — How Do People Organize Their Desks? (1983). There are people who file and people who pile, and a system must serve both. The tab doesn’t assume your folders are tidy.
C. Marshall — Rethinking Personal Digital Archiving (2008). “Benign neglect”: much of a personal archive is put away and never looked at again, and that’s fine. It’s why a quarter of the documents stays out of the themes without that being a flaw.
O. Bergman, S. Whittaker — The Science of Managing Our Digital Stuff (2016). People find things by navigating, not searching: structure matters more than search. It’s why a theme’s files are shown in their folders, with search below.
J. Stasko, E. Zhang — Focus+Context Display and Navigation Techniques for Enhancing Radial, Space-Filling Hierarchy Visualizations (2000). The folder tree as a disc of slices, from the center to the edge. It’s the layout of the cloud: each slice is a folder, and a scattered theme looks scattered.
B. Shneiderman — The Eyes Have It (1996). Overview first, then zoom and filter, then details on demand. It’s the order of the sidebar: at the top what applies to everything, below the options and the themes.
T. Forte — Building a Second Brain (2022). Projects, areas, resources, archives: you organize by what’s ongoing and what’s finished. The tab says of each theme whether it’s ongoing or finished, and has you archive instead of delete.