learn
/learn
Zapisuje odkrycie z sesji we wspólnej pamięci między projektami, po zatwierdzeniu szkicu przez użytkownika.
- Autor
- kernel64 ↗
- Repozytorium
- claude-shared-memory ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 03.11.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 1
- Skrypty
- 0
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 03.11.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
Co robi i kiedy
Zapisuje odkrycie z sesji we wspólnej pamięci między projektami, po zatwierdzeniu szkicu przez użytkownika.
Wyzwalacze z opisu
Argumenty
/learn [what to remember (optional: otherwise taken from the conversation)]
Oryginalny opis (EN)
Save a discovery from the current session into the shared cross-project memory (gotcha, workaround, tool, technique, optimization) so other projects can reuse it, after the user approves the draft. Use when the user says "learn", "retiens ça pour les autres projets", "sauvegarde dans la mémoire partagée", or to offer it right after solving a non-trivial problem whose solution is reusable elsewhere.
Jak działa
- 1. Gate: is it reusable elsewhere?
- 2. Write it for a stranger project
- 3. Avoid duplicates
- 4. Save and report
Podgląd SKILL.md
---
name: learn
description: Save a discovery from the current session into the shared cross-project memory (gotcha, workaround, tool, technique, optimization) so other projects can reuse it, after the user approves the draft. Use when the user says "learn", "retiens ça pour les autres projets", "sauvegarde dans la mémoire partagée", or to offer it right after solving a non-trivial problem whose solution is reusable elsewhere.
argument-hint: "[what to remember (optional: otherwise taken from the conversation)]"
---
# Save to shared memory
The user wants to keep something learned here so their other projects benefit. What to save: $ARGUMENTS (if empty, take the most recent reusable finding from the conversation; if several, list them and save each one separately).
**Nothing is shared without the user's yes.** If the user did not explicitly ask to save, show the draft (title, summary, stack, confidence) and ask first. Claude Code also asks the user to confirm each `memory_add` and each content-changing `memory_update` (plugin hook); if they decline, leave it in the project's own memory.
## 1. Gate: is it reusable elsewhere?
Ask yourself: "Would this save time in a *different* project?"
- Yes: a gotcha + its fix, a workaround, a tool and how to run it, a measured optimization, a platform limit, a reliable technique, a good default config.
- No: this repo's decisions, its status, its file paths, its business rules. Those belong in the project's own memory. Say so in one line and stop.
- A user preference about how to work (language, commit style...) belongs in `~/.claude/CLAUDE.md`: propose that edit instead.
## 2. Write it for a stranger project
- **Title**: specific and searchable ("Nginx strips basePath when proxying Next.js /admin", not "proxy issue").
- **Summary**: one line with the key fact or fix.
- **Body**, in English, with these sections:
- `## Context`: when it applies, symptoms, the exact error message.
- `## Solution`: concrete steps, commands, config, versions.
- `## Why`: root cause or evidence (measurement, doc link).
- `## Limits` (optional): where it does not apply, what was not tested.
- Generalize: replace project names, hosts, IPs, users, emails and paths specific to one repo with placeholders (`<APP_DIR>`, `<CONSOLE_IP>`). Tool install paths on this machine are fine (they help other projects).
- **Never** include passwords, tokens, keys, connection strings. The tool refuses them anyway.
- `type`: gotcha | technique | tool | optimization | snippet | reference.
- `stack`: the platforms it applies to (e.g. `xbox360`, `flutter`, `nestjs`, `esp32`, `android`, `windows`). Empty only if truly general.
- `tags`: lowercase tool/lib/error keywords.
- `source`: the current project folder name (it is the default when omitted).
- `confidence`, honestly:
- `assumed`: believed or read somewhere, not checked here;
- `observed` (default): seen happen in this session, cause not proven;
- `tested`: you reproduced, fixed and re-checked it, or measured it, in this session;
- `user-confirmed`: only if the user explicitly confirmed it.
- `review_after` (optional, YYYY-MM-DD): for temporary workarounds or version-specific facts that may stop being true.
- Keep it short: one finding per entry, no logs, no conversation dumps, no debugging output. Quote only the exact error line that makes it searchable. Limits enforced by the tool: title 200 characters, summary 500, body 4,000.
## 3. Avoid duplicates
Run `memory_search` with the key words first. If an entry already covers it, `memory_get` it and `memory_update` it (`append` for a new finding, `from_project` = current folder, `confirm=true` if this session checked it again, a higher `confidence` if it is now tested) instead of adding. If the new knowledge replaces an entry that turned out wrong or obsolete, add the new one, then mark the old one with `memory_update(id=<old>, superseded_by=<new>)` after asking the user: nothing is deleted.
## 4. Save and report
Once the user agreed, call `memory_add` (Claude Code shows them a confirmation; if the call is refused, do not retry: say it stays local). If it answers "similar entries already exist", read them and decide: update one, or `force=true` if really different. Then tell the user in one line what was saved (id + title).
Wersje
| Repozytorium | Wynik | Licencja |
|---|---|---|
| Ta strona · kernel64/claude-shared-memory | 90 | MIT |
| lucamartins/ouroboros | 100 | MIT |
| boykush/scraps | 100 | MIT |
| nicsuzor/academicops | 100 | MIT |
| cookys/autopilot | 100 | MIT |
| joelhooks/PAROL6 | 100 | MIT |
| tw93/Waza | 100 | MIT |
| fainir/agent-harness | 100 | MIT |
| amit429/understanding-map | 100 | MIT |
Migawka bazy z 13.11.2026 · wersja paczki 556dd5b5c5aa (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 90harvest
Przenosi wiedzę z pamięci projektów Claude Code do wspólnej pamięci, każdy wpis po zgodzie użytkownika. - 100maintain
Przegląda wspólną pamięć projektów pod kątem duplikatów, sprzeczności i wpisów do sprawdzenia, a poprawia je po zgodzie użytkownika. - 90yggdrasil-memory
Przywołuje wcześniejsze decyzje, rozwiązania i wnioski z poprzednich sesji i projektów, a po rozwiązaniu problemu zapisuje je na przyszłość. - 100self-learn
Wyciąga powtarzalne wzorce z ostatnich sesji i proponuje ulepszenia zasad projektu, po zatwierdzeniu je zapisuje. - 100continuous-learning
Zapisuje wiedzę wielokrotnego użytku i automatycznie tworzy nowe skille w trakcie sesji kodowania, rozwiązywania problemów i retrospektyw. - 95Ltm
Zapisuje i przywołuje długoterminową pamięć OpenLTM: zapamiętuje fakty, sprawdza dawne decyzje i ustawia cel projektu.