wordspace

Markdown notes for Mac and iPhone, a personal wiki that grows on its own, and all the intelligence on your device.

On the iPhone you capture, on the Mac you think. What you catch on the fly waits in the Inbox until you decide; then it becomes a note, and a personal wiki grows from your notes by itself. No account, no servers of ours: the app collects no data, and every model runs on your Mac, Apple Intelligence or a small open model you download once. Six languages, free.

Website · Mac App Store · The idea and its sources · Privacy

Our models

Small, specialised, open. Each one does one job of the app, and does it on the Mac.

Built for the LLM wiki

wordspace follows the LLM wiki pattern proposed by Andrej Karpathy (2026): your notes are compiled once into linked topic pages that a model keeps up to date, instead of being searched again at every question. The notes stay the source of truth, the pages stay readable Markdown, and every page cites the notes it comes from.

Keeping such a wiki honest is a stream of small decisions on every new note. We split the work so that each step is done by the cheapest thing that does it well, a cascade (Viola and Jones 2001; FrugalGPT, Chen et al. 2023):

1. Code decides what is decidable. Two different numbers or dates for the same thing are a contradiction, and comparing them needs no neural network: on 1,000 measured cases the deterministic check has precision 1.00.
2. The decider answers the closed questions the code cannot: do these two notes contradict each other, is this value decided or only proposed, does a later note update an earlier one or do two sources disagree, does this note belong on this page. One forward pass, about a tenth of a second.
3. A generative model writes: the pages, the quotes and the reason behind a contradiction, and judges only the few cases the decider leaves uncertain.

On the app's 154 hand-written gold cases, the contradiction check with our decider raises 2 false alarms instead of 7 with LiquidAI's d1-3B, and calls the generative model a third less often. On its own, the decider answers 280 of 292 test questions, against 270 for d1-3B. Neither difference is statistically significant yet, and we say so on every model card.

How we build them

• Open bases only: Apache 2.0 or MIT, with the licence re-read from the downloaded files, and a NOTICE that lists every change we made.
• No personal data, ever. Training data is synthetic, generated from sentences we write by hand in six languages: Italian, English, Spanish, French, German and Portuguese.
• Tests only measure. The test sets are never used to write training data, and any generated case that shares eight consecutive words with a test is thrown away.
• Against shortcuts: every pattern comes with its minimal pair, the same sentence that must get the other answer, so the model learns the decision and not a surface cue (Gardner et al. 2020; McCoy et al. 2019).
• Measured honestly: always against the strongest model we know, on the same cases, with Wilson intervals and exact McNemar tests, and the losses reported next to the wins.
• Small and quantized with care: MLX, quantized only as far as accuracy holds. The decider stays at 6 bits because at 4 bits it loses the cases that matter.