Przejdź do treści
Biblioteka skilli
AI i agenci · Automatyzacja

update-mem

/update-mem

Zapisuje notatkę sesji i stan pamięci w repozytorium Markdown, przed kompaktowaniem, na koniec sesji i po poprawkach.

Repozytorium
second-nature ↗
Licencja
MIT
Ostatnia zmiana repozytorium
12.11.2026

Ocena i bezpieczeństwo

100/ 100Klasa: dobry
Format25 / 25
Frontmatter z nazwą i opisem, poprawna nazwa, opis do 1024 znaków, nazwa zgodna z folderem.
Odkrywalność25 / 25
Opis mówi, kiedy użyć skilla, ma sensowną długość i jest w 3. osobie.
Treść35 / 35
Co najmniej 150 słów i do 500 linii, nagłówki, listy kroków, przykłady albo kod.
Higiena15 / 15
Bez zaślepek typu TODO, bez ścieżek Windows, z licencją.
Ryzykobrak sygnałów w skanie
Pliki w paczce
1
Skrypty
0
Komentarze HTML
0
Adresy URL w SKILL.md
1
Przestarzałość

Brak sygnałów. Repozytorium zmienione 12.11.2026, bez odwołań do wycofanych modeli i API.

Licencja

MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.

Co robi i kiedy

Zapisuje notatkę sesji i stan pamięci w repozytorium Markdown, przed kompaktowaniem, na koniec sesji i po poprawkach.

Wyzwalacze z opisu

update memsave itsave thissave karremember thisyaad rakhnote karoupdate vault
Oryginalny opis (EN)

Save to memory. Use when the user says "update mem", "save it", "save this", "save kar", "remember this", "yaad rakh", "note karo", "update vault", "save session", or "flush memory"; and without being asked BEFORE any compact, at session end, when context grows long, or right after the user corrects a mistake. Writes the session note and hot.md in the markdown memory vault and routes each lesson to a code guard, test, hook, or single rule line. Any project, any CLI agent.

Jak działa

  1. 1. Collect (this session only, concrete items only)
  2. 2. Each lesson: highest rung that can hold it
  3. 3. Route prose by scope: ONE home, never two
  4. 4. Session note (state)
  5. 5. hot.md
  6. 6. Budgets must pass
  7. 7. Reply
Sekcje z nagłówków SKILL.md, w oryginale · 80 linii, 758 słów
Podgląd SKILL.md
---
name: update-mem
description: Save to memory. Use when the user says "update mem", "save it", "save this", "save kar", "remember this", "yaad rakh", "note karo", "update vault", "save session", or "flush memory"; and without being asked BEFORE any compact, at session end, when context grows long, or right after the user corrects a mistake. Writes the session note and hot.md in the markdown memory vault and routes each lesson to a code guard, test, hook, or single rule line. Any project, any CLI agent.
---

# update-mem — lessons become guards, state goes to the vault

Vault location: `~/.agents/memory` (override via `$AGENT_MEMORY_VAULT`). Protocol: `MEMORY-PROTOCOL.md` at the vault root.
Setup and architecture: https://github.com/anki1kr/second-nature/blob/main/docs/memory-system.md. Utility scripts live in `~/.agents/memory/bin/` once the vault is set up (`install.sh --memory`).

**Never block.** This often runs seconds before a compaction. Do not ask the user questions here; write down any ambiguity as an open item. Work silently.

## 1. Collect (this session only, concrete items only)

- **Lessons:** each user correction, repeated mistake, new rule or preference, new bug pattern, faster workflow.
- **State:** what changed (exact file paths), what was verified and HOW, what is NOT verified, what remains open.

Nothing new since the last flush → reply `✓ Nothing new to save.` and stop.

## 2. Each lesson: highest rung that can hold it

| Rung | Form | Costs tokens every turn? |
| --- | --- | --- |
| 1 | Guard in the product's own code (refuse, validate, filter) | No |
| 2 | Automated test that fails when the rule is broken | No |
| 3 | Hook (git pre-push, CI gate, harness hook) or platform linter rule | No |
| 4 | ONE line of prose in a standing rule file | Yes |

If Rung 1 to 3 is possible, small (~20 lines), and you are already working in that repository → implement it, run it once, and record it in the session note. Otherwise write the Rung 4 line AND an open item "promote to rung N". When running immediately before a compaction → record only, implement nothing.

## 3. Route prose by scope: ONE home, never two

| What it is | Home | Form |
| --- | --- | --- |
| How to work in THIS project | `~/.agents/projects/<repo>/core.md` (hard rule) or repo `AGENTS.md` | One line + pointer; narrative belongs in project docs |
| Behaviour that holds across all projects | `wiki/meta/lessons-learned.md` | One line: rule, why, trigger; bump `updated:` |
| A gotcha of one domain (shell, browser QA, CSS, agents, SEO) | `wiki/meta/rules/<domain>.md` | One line; domain reference lives here |
| User interaction preferences across every tool | `~/.agents/AGENTS.md` | One line, strictly within budget |
| Project state and history (not standing rules) | The session note, linked from `wiki/projects/<hub>.md` | Verified facts only |

- Unsure whether a rule is project-specific or general → choose project (smaller blast radius).
- If the repository has no `~/.agents/projects/<repo>/` → run `python ~/.agents/memory/bin/project-init.py <repo>` first (see `../../memory/bin/project-init.py`).
- A harness-specific memory folder (for example `~/.claude/projects/*/memory`) is a local cache. Write there only what that tool requires, never as the sole copy.
- Never duplicate a repository fact inside the vault: link directly to the repo file that holds it.

## 4. Session note (state)

File path: `wiki/sessions/YYYY-MM-DD-<topic-slug>.md`. If a note for today's session already exists → update it; never create a duplicate.

- **Frontmatter:** `type: session`, `title`, `date` (ISO-8601), `project: "[[projects/<hub>]]"`, `prev: "[[<previous note>]]"`, `tags`.
- **Body (facts only):**
  - **Done:** exact touched paths.
  - **Verified:** the command or check run, and its result.
  - **Not verified:** areas touched but unverified.
  - **Guards added:** rung level and target file.
  - **Open:** pending work, including lessons awaiting promotion.

## 5. hot.md

One line at the top of the `## Recent sessions` section:
`[[<note>]] — **<headline>**: summary of progress + key open items.`
Keep it under 520 characters; detailed narrative belongs in the session note. If an entry for this session already exists, update it in place.

## 6. Budgets must pass

Check budgets and rotate entries:
```bash
python ~/.agents/memory/bin/mem-budget.py --fix "<repo>/AGENTS.md" 12000
```
Then regenerate vault indexes:
```bash
python ~/.agents/memory/bin/mem_index.py
```
(Regenerates `wiki/index.md`, `projects/_index.md`, `rules/_index.md`, and the recent-sessions block; never edit those by hand).
Omit the last two arguments to `mem-budget.py` if the repository lacks a dedicated agent instruction file.
`--fix` rotates older `hot.md` lines beyond the cap into `hot-archive.md`. The script exits with code 1 if `hot.md`, `lessons-learned.md`, `AGENTS.md`, or the project file exceeds its size limit.
Exit code 1 → trim what you just wrote, or move historical narrative into `docs/`. Never increase budget limits or bypass this check.

## 7. Reply

`✓ Saved to vault.`
Add at most ONE follow-up line only if something requires user intervention: a budget limit that could not be resolved, a rule without a clear home, or a high-priority lesson awaiting promotion.
Use absolute dates. Never output verbatim secret quotes, and never execute `git commit` or `git push` unless explicitly asked.

Wersje

W bazie nie ma innych skilli o nazwie update-mem.

Migawka bazy z 13.11.2026 · wersja paczki 34a0af12b00b (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

  • 100mem-search
    Przeszukuje pamięć Claude-Mem (cmem.ai) użytkownika, by odtworzyć kontekst z poprzednich sesji o projekcie, decyzji, pliku lub zadaniu.
  • 100m0-handoff
    Zapisuje lub odczytuje compact_checkpoint przed kompaktowaniem i na koniec sesji, żeby zimny start zamienił się w kontynuację także w Cursorze lub Claude Code.
  • 95wicked-garden-mem-capture
    Na koniec sesji wyłapuje trwałe wnioski z rozmowy i zapisuje je do pamięci, a reguły trafiają do przeglądu.
  • 100mem-forever
    Zapisuje pamięć agenta w jednym repozytorium Git: profil, pliki wspomnień, notatkę przekazania i rejestr kluczy, a także wczytuje ją na starcie sesji.
  • 100cowork-mem
    Zapewnia trwałą pamięć między sesjami Cowork: przypomina, co działo się wcześniej, i zapisuje decyzje, zmiany plików, wnioski, błędy i użycie narzędzi.
  • 100memory-update
    Skanuje transkrypty sesji w poszukiwaniu sygnałów wartych zapisu (poprawki, preferencje, decyzje, odwołania) oraz zapisuje pliki auto-pamięci z poprawnym frontmatter i wpisem w MEMORY.md.

Uwagi i zgłoszenia

Wiadomość trafi do Bartosza Sokolińskiego. Adres e-mail jest potrzebny tylko do odpowiedzi. Szczegóły w polityce prywatności.