Blog of Justin Craig-Kuhn

StoryKey is a proof of concept Brainwallet inspired by BIP39 written in Ruby. It converts an arbitrary string of data, such as a cryptocurrency private key, into an English paragraph intended for longterm human memory. It also assists in decoding the story back into its original form. Optionally, a visual representation of the paragraph is also provided using OpenAI DALL-E.

CLI Screenshot of the StoryKey command line interface

I picked up a hardware to take control of my own crypto keys. Hardware wallets use an initial private seed to generate identities for many common cryptocurrencies. The private seed is recoverable from a seed phrase, which consists of 12 or 24 words that the user must memorize and keep private. This enables recovery of all derivative identities, and therefore funds, in case of hardware loss or failure. I found the process of memorizing these words to be challenging. The lexicon is somewhat random, with different parts of speech mixed in randomly.

I theorized that the phrase might be easier to remember if it used different parts of speech grammatically to form images in one’s mind. These images could then be visually presented as a mnemonic device along with the text. Perhaps some people may find remembering the images easier than the text, or vice versa. StoryKey uses DALL-e to optionally generate an image depicting the story in a set of panels. Unfortunately there are restrictions around fictional characters and celebrities so the images don’t adequately reflect Marge Simpson or Jimi Hendrix, for example. Perhaps in the future generative image models will have relaxed rules around this.

Dalle An image produced by StoryKey representing a private key

It would take some kind of survey or test administered to actual end users to find out if the software produces any significant change in memorability. I may try to conduct that at some point, but for now I will continue to tinker with the idea. Image generation AI’s continue to improve, and an important next step will be using more detailed prompts to produce consistent memorable imagery. I may also be able to use large language models to produce better lexicons or even culturally specific lexicons.

Takeaways:

  • It is difficult to produce lexicons appropriate for the widest audience - cultural references are useful but not universal
  • Generative image AI in 2023 produced a lot of randomness
  • It is unclear whether structured phrases are easier to memorize than random words

Updated: