Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Stores observations and friction patterns into the learning loop. Over time, observations accumulate evidence. When confidence reaches a threshold, they become eligible for `/rethink` synthesis. This is how the system learns from experience.
.claude/skills/miosa-osa-remember/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | 11% | 0% |
| case-03 | ✗→✓ | ▲ Improved | -74% | 0% |
| case-04 | ✗→✓ | ▲ Improved | -55% | 0% |
| case-05 | ✗→✓ | ▲ Improved | -34% | 0% |
| case-06 | ✗→✓ | ▲ Improved | -63% | 0% |
Capture friction patterns and observations for the learning loop.
3-mode friction capture:
Observations stored in observations table. When 3+ observations accumulate in the same category, escalates for rethink.
/remember "always check duplicates before ingesting"
/remember --contextual # Scan recent sessions
/remember --mine # Bulk extract from sessions
/remember --escalations # Show patterns ready for rethinkObservations are auto-classified into: process, people, tool, decision, pattern, friction
bashcd engine && mix optimal.remember "observation text" cd engine && mix optimal.remember --contextual cd engine && mix optimal.remember --mine cd engine && mix optimal.remember --escalations
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-02 | fail→fail | 4,869 | 3,572 | -27% | 1 | 1 | 0% | 663 | 697 | +5% | 0 | 0 | — |
case-01 | fail→pass | 5,227 | 4,782 | -9% | 1 | 1 | 0% | 795 | 883 | +11% | 0 | 0 | — |
case-03 | fail→pass | 12,541 | 4,484 | -64% | 1 | 1 | 0% | 1,811 | 479 | -74% | 0 | 0 | — |
case-04 | fail→pass | 10,053 | 2,074 | -79% | 1 | 1 | 0% | 1,300 | 583 | -55% | 0 | 0 | — |
case-05 | fail→pass | 5,364 | 1,766 | -67% | 1 | 1 | 0% | 719 | 472 | -34% | 0 | 0 | — |
case-06 | fail→pass | 23,262 | 5,678 | -76% | 1 | 1 | 0% | 1,671 | 616 | -63% | 0 | 0 | — |
case-07 | fail→pass | 15,745 | 2,233 | -86% | 1 | 1 | 0% | 893 | 570 | -36% | 0 | 0 | — |
case-08 | fail→pass | 11,237 | 3,254 | -71% | 1 | 1 | 0% | 1,663 | 682 | -59% | 0 | 0 | — |
case-09 | fail→pass | 17,618 | 1,634 | -91% | 1 | 1 | 0% | 2,715 | 515 | -81% | 0 | 0 | — |
case-10 | fail→pass | 11,039 | 3,319 | -70% | 1 | 1 | 0% | 1,665 | 775 | -53% | 0 | 0 | — |
case-11 | fail→pass | 8,190 | 3,212 | -61% | 1 | 1 | 0% | 1,230 | 690 | -44% | 0 | 0 | — |
case-12 | fail→pass | 12,746 | 1,846 | -86% | 1 | 1 | 0% | 2,017 | 542 | -73% | 0 | 0 | — |
case-13 | pass→pass | 10,373 | 5,556 | -46% | 1 | 1 | 0% | 1,298 | 1,014 | -22% | 0 | 0 | — |
case-14 | pass→pass | 11,580 | 2,791 | -76% | 1 | 1 | 0% | 1,610 | 662 | -59% | 0 | 0 | — |
case-15 | pass→pass | 7,631 | 2,536 | -67% | 1 | 1 | 0% | 1,039 | 601 | -42% | 0 | 0 | — |
case-16 | pass→pass | 15,077 | 1,629 | -89% | 1 | 1 | 0% | 2,243 | 418 | -81% | 0 | 0 | — |
case-17 | pass→pass | 6,574 | 2,350 | -64% | 1 | 1 | 0% | 1,036 | 531 | -49% | 0 | 0 | — |
case-18 | fail→pass | 10,057 | 2,200 | -78% | 1 | 1 | 0% | 1,364 | 614 | -55% | 0 | 0 | — |
case-19 | fail→pass | 12,712 | 2,234 | -82% | 1 | 1 | 0% | 1,627 | 586 | -64% | 0 | 0 | — |
case-20 | pass→fail | 3,401 | 6,742 | +98% | 1 | 1 | 0% | 444 | 716 | +61% | 0 | 0 | — |
case-21 | pass→pass | 12,823 | 10,178 | -21% | 1 | 1 | 0% | 1,942 | 2,174 | +12% | 0 | 0 | — |
case-22 | pass→pass | 7,963 | 4,085 | -49% | 1 | 1 | 0% | 1,180 | 855 | -28% | 0 | 0 | — |
case-23 | pass→pass | 12,689 | 9,933 | -22% | 1 | 1 | 0% | 1,990 | 1,577 | -21% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 23 cases were attempted. The headline lift of +52 percentage points is the difference between those two pass rates over the 23 comparable cases. 1 case got worse with the skill loaded, and it is included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.