AI notes should show their receipts.

Personal notes contain high-consequence trivia

The things buried in notes look small until the moment they matter: the cancellation deadline, the quote that included one part but excluded another, the exact model number, the promise made on a call.

A fluent answer can erase useful uncertainty. “The plumber quoted £185” sounds complete, while the source note may add that damage behind the tiles is excluded. Compression helps us move quickly; it can also remove the clause that changes the decision.

A receipt has three jobs

These jobs are more important than making the answer sound polished. A short answer with visible evidence is often more useful than a comprehensive answer that hides its basis.

• Verification: let the person confirm the wording and surrounding context.
• Recovery: reopen the original note rather than trapping the answer in chat.
• Calibration: reveal whether the answer rests on one clear source, several mixed sources, or weak evidence.

What a responsible answer should look like

Lead with the smallest useful answer. Attach the source notes that materially support it. Preserve dates and qualifiers. If notes disagree, show the disagreement instead of silently choosing a winner. If nothing supports the answer, say so.

This pattern also changes user behaviour. People can ask rougher, more natural questions because they know they can inspect what came back.

The standard we want

AI memory tools should be judged not only by whether they produce the right sentence, but by whether they help a person retrace the path to that sentence. The benchmark in our Lab is intentionally simple: every answer names and opens its source.

Answers are convenient. Receipts make them dependable.