Skip to content
Memorise
Back to the site

Writing

What we measured

Notes from building a voice companion for people who live alone. Every number in here came off our own data, including the ones we would rather not have found.

Ranked the way Ellie ranks a memory
0.60

How well it matches what you asked for

0.25

How load-bearing it is, and how badly it ages

0.15

How recently it was written

The weights, the formula and the importance-scaled decay are the real ones. The one difference: Ellie compares meaning using 1536-dimension embeddings, so “the singer of that song” finds the right memory without sharing a word with it. Your browser has no embedding model, so this counts shared words instead.

6 posts, in her order

  1. 5 minThe absence of a notification means nothingHow to write down the limits of a feature that tells a family somebody may have fallen — before the first time it fails to.safetytrustwriting
  2. 4 minThe voiceprint nobody asked forA privacy control that works by collecting biometric data, built to be invisible — which is precisely the property the law prohibits.privacybiometricstrust
  3. 4 minNobody opens the conversationAlmost nine in ten conversations with our companion contained nothing from the person she was talking to. The fix was not a better prompt.lonelinessproductmeasurement
  4. 3 minWe paid to learn nothing, 86% of the timeA language-model call on every single turn, with a 14% hit rate. What it cost, and why the fix was a batch of one.costarchitecturemeasurement
  5. 4 minHebrew has no italicsEight languages, two of them right-to-left, and the small things that only break for somebody who actually reads the script.languagesdesignaccessibility
  6. 3 minThe gradient that was unreadable in both directionsA day-to-night gradient behind text looks beautiful and fails contrast for both text colours at once. Measured: 3.85:1 at its worst.designaccessibilitymeasurement