agentsclimarketplace

Experience memory

Skill alidevai/chiron_agent/.claude/skills/experience-memory

Autonomous, self-improving AI agent platform for Claude Code — safe capability & skill acquisition, MCP/tool discovery, sandboxed policy gates, and token-efficient learning.

Install
npx -y skills add alidevai/chiron_agent --skill experience-memory

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

3 things to look at

  • 20 days oldThe repository was created 20 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
  • no licenseNo license file was found in the repository. Code published without one is not open source by default, so using it at work is a question for whoever answers licensing questions where you are.
  • 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Token-verimli kendi kendine ogrenme. Goreve baslarken gecmis dersleri leksikal olarak geri cagirir (yalnizca en ilgili 3-5 kisa ders enjekte edilir); gorev bitince genellenebilir, tekrar eden bir ders cikarsa kaydeder. Amac, transkriptleri context'e doldurmadan deneyimden ogrenmektir.

SKILL.md

2.8 KB, as published. Nobody here has run it

Experience Memory (token-verimli ogrenme)

Bu skill, sistemin deneyimden ogrenmesini saglar AMA token patlamasina yol acmaz. Butun sir sudur: ogrenme deterministik Python'da saklanir; LLM'e yalnizca kucuk, kirpik dersler enjekte edilir.

Neden token-verimli?

  • Depolama ve geri cagirma python -m core learn (SQLite) ile yapilir -> LLM token'i 0.
  • Bir ders ~1-2 satirdir; transkript, uzun aciklama, kod bloklari SAKLANMAZ.
  • Geri cagirma leksikaldir (kelime ortusmesi); embedding API'si yok, ek maliyet yok.
  • Goreve en fazla 5 ilgili ders enjekte edilir (progressive disclosure). Tum ders defteri asla context'e yuklenmez.
  • Anti-bloat: benzer ders tekrar eklenince yeni satir acilmaz, mevcut dersin sayaci artar. Boylece defter kucuk kalir ve her gorevde az token tuketir.

Gorev BASINDA (recall)

Gorevin alanini ve anahtar kelimelerini belirle, sonra:

python -m core learn recall --domain <alan> --query "anahtar kelimeler" --k 5

Donen 0-5 dersi kisaca dikkate al. Ders yoksa hicbir sey enjekte etme (sifir maliyet). Bunu capability-gap-analysis ile ayni anda yap (gorev acilis ritueli).

Gorev SONUNDA (distill) — ama her zaman degil

Ders eklemeden once ponytail sorusunu sor (bkz. [[minimalist-engineering]]): "Bu gercekten yeni, genel ve tekrar edecek bir ders mi?" Degilse ekleme.

Bir ders yalnizca su kriterlerin HEPSI saglaninca eklenir:

  • Ayni/benzer gorev tekrar edebilir
  • Prosedur genel ve tasinabilir (projeye/sirra ozel degil)
  • Basari veya basarisizlik acik bir ogrenim veriyor
  • Kullaniciya ozel sir/veri icermiyor

Eklerken kisa tut:

python -m core learn add "KURAL (1-2 cumle)" --domain <alan> \
    --trigger "ne zaman gecerli" --why "gerekce"

Pekistirme ve terfi

  • Bir ders bir gorevde ise yaradiysa/yaramadiysa: python -m core learn reinforce <id> win|loss
  • Bir ders yeterince tekrar edip (uses >= 3) net pozitifse skill adayidir: python -m core learn promotion-candidates Bu adaylari skill-creator-safe ile tam skill'e donustur. Boylece defter degil, yalnizca KANITLANMIS tekrar eden desenler kalici skill olur (context sismesini onler).
  • Periyodik temizlik: python -m core learn prune (kullanilmayan dersleri budar).

Kullanilmamasi gereken durum

  • Tek seferlik, onemsiz gorevlerde ders ekleme; defteri gurultuye bogar.
  • Uzun metin/transkript kaydetme; ders tek satirlik kural olmalidir.

Kaynaklar

  • Motor: core/learning.py
  • Ilgili: [[minimalist-engineering]], [[capability-gap-analysis]], [[skill-creator-safe]]

Keep looking

Skills are one crate of 328,083. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.