An open test kit for AI journal memory and voice

Download the test kit

Version 1 includes a fictional dated diary, answer criteria, repeat-run instructions, a voice reading script and a blank results record. The fixtures are in English so each run can use identical input. No member journals are included.

Test memory with checkable facts

1. Start a fresh test journal. Add the fictional entries in date order and allow the app’s normal indexing to finish. Record the app, plan, version, date and any feature restrictions.
2. Ask each question in a fresh conversation three times. Keep the full answers, retrieved evidence when available, and any refusals or errors. Do not retry until you get a favorable answer.
3. Check correct facts, later corrections, two different people with the same first name, and a question whose answer is absent. Score the supplied criteria separately. Missing evidence is not permission to invent an answer.
4. Report each run and its limitations. A test that supplies all passages directly measures answering, not retrieval. A seeded history does not establish months of real-world memory or prove that deleted information is erased.

Test transcription without guessing

Read the supplied fictional script into each app. Use the same consented recording where upload is supported, or document the microphone and playback setup. Synthetic speech must be labeled as synthetic.

1. Keep the exact reference script and raw transcript. Count substitutions, deletions and insertions, then divide their total by the reference word count. Keep the normalization rules with the result.
2. Check names, dates, numbers and negation separately. A low overall word error rate can still conceal a changed meaning.
3. Record whether the original audio can be replayed, corrected, downloaded or deleted, and which plan was required. One speaker or one language cannot establish general accuracy.

Dayora: measured on 20 September 2026

The shipped answer and transcription functions ran locally on the published fixtures. Factual checks: 18 of 18 passed across six questions and three repeats. Voice: zero word errors in each of three runs on the same 57-word synthetic recording, after the published normalization. These are publisher-reviewed results, with Codex assistance, not an independent ranking.

All ten entries were retrieved for every question. This tests answer grounding on a small history, not large-library retrieval or months of memory. Missing-fact answers correctly declined to invent a name but included unnecessary unrelated details. Citation precision and concision were not scored. Production Ask also requires Plus and rollout access.

Mindsera and Rosebud: no hands-on result yet. Their documented features are compared separately; neither has been assigned a benchmark score.

Results and run status

The downloadable results record identifies completed runs, raw outputs and untested products. Only completed tests count toward a result; vendor documentation is not a measured score.

What these tests cannot tell you

These small synthetic tests do not measure emotional benefit, clinical safety, every language, every accent or long-term retention. AI responses vary. Dayora has a commercial interest in the outcome; publish failures and replicate the work before relying on a comparison. The kit and its original fixtures are available under CC0.