Files

T

Antigravity Agent 504fd5fb42 [G1-Sync] Manual knowledge update

2026-05-10 22:08:15 +09:00

4.5 KiB

Raw Blame History

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, tech_stack

title

P-Reinforce

매 한 줄

"매 wiki entry 의 자동 강화 marker". P-Reinforce 는 Claude Code agent 가 wiki node 를 재방문할 때마다 frontmatter 의 last_reinforced 를 갱신하고 missing graph link 을 보충하는 light-weight reinforcement loop. 매 hand-curated knowledge graph 의 staleness 방지.

매 핵심

매 trigger 조건

agent 가 file 을 Read 하고 substantive change (>= 3 lines) 을 적용한 경우.
backlink resolver 가 dangling [[X]] 을 발견 후 stub 을 만든 경우.
weekly cron (/loop 7d /reinforce) 의 sweep.

매 frontmatter 의 강화 field

last_reinforced — ISO date.
confidence_score — 0..1, decay 0.02/month, +0.05 per reinforcement.
verification_status — applied | pending | redirected.
source_trust_level — A/B/C.

매 응용

Wiki staleness audit — last_reinforced < 90d 의 file 을 batch 로 재방문.
Confidence-weighted retrieval — RAG 가 high-confidence node 을 prefer.
Graph hygiene — orphan / dangling link 을 reinforcement pass 에서 fix.

💻 패턴

Frontmatter reinforcement (Python)

import re, yaml, datetime, pathlib

def reinforce(path: pathlib.Path, bump: float = 0.05):
    text = path.read_text()
    m = re.match(r"^---\n(.*?)\n---\n(.*)$", text, re.DOTALL)
    if not m:
        return False
    fm = yaml.safe_load(m.group(1)) or {}
    fm["last_reinforced"] = datetime.date.today().isoformat()
    fm["confidence_score"] = min(0.99, fm.get("confidence_score", 0.5) + bump)
    fm["verification_status"] = "applied"
    new = "---\n" + yaml.safe_dump(fm, sort_keys=False, allow_unicode=True) + "---\n" + m.group(2)
    path.write_text(new)
    return True

Decay sweep (cron)

from datetime import date, timedelta

def decayed(fm: dict, days: int = 90) -> bool:
    last = date.fromisoformat(fm.get("last_reinforced", "1970-01-01"))
    return (date.today() - last) > timedelta(days=days)

stale = [p for p in pathlib.Path("10_Wiki").rglob("*.md")
         if decayed(load_fm(p))]

Dangling link harvest

import re
LINK = re.compile(r"\[\[([^\]|#]+)")
def dangling(root):
    existing = {p.stem for p in root.rglob("*.md")}
    for md in root.rglob("*.md"):
        for tgt in LINK.findall(md.read_text()):
            if tgt not in existing:
                yield md, tgt

Reinforcement-aware retrieval (BM25 + confidence)

def score(doc, query):
    bm = bm25(doc.text, query)
    conf = doc.frontmatter.get("confidence_score", 0.5)
    age = days_since(doc.frontmatter["last_reinforced"])
    decay = max(0.5, 1 - age / 365)
    return bm * conf * decay

Claude Code hook (settings.json)

{
  "hooks": {
    "Stop": [{
      "command": "python ~/scripts/reinforce_touched.py"
    }]
  }
}

매 결정 기준

상황	Approach
매 active editing	inline reinforcement on Write
매 background hygiene	weekly cron sweep
매 large refactor	manual `verification_status: applied` bump
매 redirect/duplicate	`verification_status: redirected`, no decay

기본값: Stop hook + weekly sweep 의 조합.

🔗 Graph

부모: Knowledge-Graph · Wiki-Architecture
변형: Auto-Curation · Spaced-Repetition
응용: RAG · Agent-Memory
Adjacent: Obsidian-Frontmatter · Confidence-Scoring

🤖 LLM 활용

언제: 매 long-lived wiki / second-brain 의 staleness 가 문제일 때. 언제 X: 매 ephemeral note / scratchpad — overhead > benefit.

❌ 안티패턴

Reinforce without read: blind date bump — confidence 의 의미를 잃음.
Unbounded confidence growth: cap 없이 +0.05 누적 — 매 1.0 saturation. cap 0.99.
Decay-only no-bump: stale 만 잡고 fresh 강화는 없음 — graph 가 균질하게 늙음.

🧪 검증 / 중복

Verified (in-house spec, CLEANUP_SPEC.md).
신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — full content for P-Reinforce framework

4.5 KiB Raw Blame History