Preset questions
Instant, pre-computed from the full dataset — no AI calls, no waiting.
Or ask your own
Do not paste personal information. This static demo never sends anything anywhere — every answer is computed in your browser from the bundled data.
Answer
Pick a preset or ask your own question — the answer, the tables behind it, and the raw evidence episodes appear here.
Where it breaks — failure gates
Episodes with a gate · main chart excludes not_search_related and unclear (reported separately)
By era
What people remember vs what they forget
Share of all remembered / missing cues by cue type (T2 cue matrix)
How they search
Query style across episodes
What they do after it fails
Workaround after a failed retrieval
Figures from the offline pipeline
Generated by code (analyze.py), click to open full size
Every episode, searchable
Filter and search the raw evidence
Try it: paste a vague photo memory
The engine finds the 8 nearest real episodes and estimates the likely failure point — the same nearest-episode method the full app uses.
Nearest-episode estimate (no live AI) — majority vote over the 8 nearest episodes by TF-IDF similarity, computed in your browser. Do not paste personal information — nothing leaves this page.
Opportunity scorecard
photo_type × failure_gate cells with n≥8 · drag the weights and watch the ranking move live
score = 100·(w_f·freq/maxfreq + w_s·sev + w_g·gap + w_l·lev). p_top3 = probability of a top-3 rank over 1,000 Dirichlet(20·[0.35, 0.25, 0.20, 0.20]) weight draws, computed offline. Only 3 cells clear n≥8 — treat the ranking as directional.
Limitations of this evidence
- App-store and Play reviews skew to complaints; Reddit skews to Android power users.
- The Reddit sample covers a 2026-03-19 → 2026-10-06 archive window (rate-limit deviation, stated on the deck).
- Gates are model-inferred. Human agreement on the held-out test set (n=40, prompt v1): is_retrieval 97.5%, photo_type 100%, failure_gate 67.5% (κ=0.569), outcome 80.0%.
- Results show failure types in a convenience sample — not prevalence in the Google Photos population.
- 182 of 765 screened relevant documents became extraction episodes; only 3 scorecard cells reach n≥8.
Discover more: fresh public feedback
Search live feedback about the app from the Google Play Store, the App Store, and Reddit — then get a Gemini analysis when AI is available, or a code-only reading when it isn't.
Documents are fetched live from public sources for your query, then analyzed by Gemini when available. Do not type personal information — your query is sent only to this site's search service.