The idea in practice
A fair memory test includes recall, correction, forgetting and relevance. Use equivalent fictional scenarios across tools and record failures rather than selecting one impressive demonstration.
Try the distinction on a real task
This is an illustrative request, not a customer result or a promise that every connected service supports every step.
Run this fictional preference task, correct one constraint, and explain the output against the latest instruction.
How to apply it
- Keep test identities separate
- Use identical goals and evidence
- Record both successes and failures
What to check
Check whether the output respected the correction and avoided unrelated private context.
A mistake to avoid
Do not claim a universal winner from one prompt, one user account or different amounts of setup.
Using Bris for this work
Bris is designed to carry useful preferences into later work. Tell it what should stay relevant, correct details when they change, and ask what it remembers before relying on an important constraint. A remembered preference is context; it is not permission to send, book or buy.