메뉴
Give your AI long-term and short-term memory: keyword-searchable session history (FTS5 index) plus a hot cache of current project state loaded at session start. Use when AI should remember past sessions and know project state without re-explaining.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
| name | ai-memory |
| description | Give your AI long-term and short-term memory: keyword-searchable session history (FTS5 index) plus a hot cache of current project state loaded at session start. Use when AI should remember past sessions and know project state without re-explaining. |
AI doesn't remember anything between sessions. Every new chat starts blank — no history, no context, no idea what you shipped yesterday.
This skill gives your AI two things it's missing:
Side 1: Can't find the past. You discussed an auth refactor two weeks ago. Made a decision. Wrote a great prompt. Now it's gone — buried in hundreds of session files with no search, no index.
Side 2: Doesn't know the present. What shipped yesterday? What's blocked? When's the deadline? AI has no idea until you paste context manually. Every session. Every morning.
อ่านไฟล์ JSONL ทุกไฟล์ใน ~/.claude/projects/ ดึงแค่ข้อความ user + assistant ออกมา สร้าง SQLite database ที่มี FTS5 full-text search index ให้ search ด้วย keyword ได้ทันที เขียนเป็น Python script ชื่อ session-store.py ที่รับ command search, ingest, stats
English version:
Read all JSONL files in ~/.claude/projects/ and extract only user + assistant messages. Create a SQLite database with FTS5 full-text search index that can search by keyword instantly. Write it as a Python script called session-store.py that accepts commands: search, ingest, stats
One prompt. AI writes the script, runs ingest, ready to search.
# Search across all sessions
python3 session-store.py search "refactor authentication"
# Output:
# [2026-05-10] interactive:a3f2... (42,318 chars)
# ...ต้อง refactor authentication flow เพราะ Privy session...
# Re-index after new sessions
python3 session-store.py ingest
# Show index stats
python3 session-store.py stats
Key components:
type: "human" and type: "assistant" messagestokenize='unicode61' for multilingual supportsearch "keyword", ingest (rebuild index), stats (show counts)No cloud, no API keys, no vector DB, no embeddings. FTS5 is built into Python's sqlite3.
อ่านไฟล์พวกนี้: [standup log, decisions.md, schedule.env, tester-gate.md] สรุปให้เหลือแค่ข้อมูลที่จำเป็น — เมื่อวาน ship อะไร, blocker ตอนนี้คืออะไร, deadline เมื่อไหร่, decision ไหนยัง active — save เป็น hot-cache.md
English version:
Read these files: [standup log, decisions.md, schedule.env, project-status.md] and summarize into only the essential information — what shipped yesterday, current blockers, upcoming deadlines, which decisions are still active — save as hot-cache.md
Replace the file list with whatever holds your project state: git log, TODO files, kanban boards, decision journals, sprint trackers, open PRs.
# Hot Cache — 2026-05-24
## Last Standup
- feature X ✅ shipped
- feature Y BLOCKED on [person]
- feature Z IN PROGRESS
## Active Decisions
- Auth approach: social-first (Privy) — active
- Deployment: Cloud Run — decided, not revisiting
## Deadlines
- Launch: June 2
- Demo: May 28
## Recently Modified Files (last 24h)
- src/auth/privy.ts
- contracts/Parimutuel.sol
0 */4 * * * cd /path/to/project && claude -p "Read [source files]. Update hot-cache.md." >> /dev/null 2>&1
This is the key step. Add these lines to your CLAUDE.md (or your AI tool's instruction file):
Before starting work:
1. Read data/hot-cache.md for current context (blockers, decisions, deadlines)
2. If you need info from a past session, use: python3 session-store.py search "keyword"
Now every session starts with context and has access to history. No manual pasting.
unicode61 tokenizer works with Thai, English, and mixed text.Turn a content creator's AI stack into a factory that sounds like THEM. One voice spine (CONTENT.md) governs every format — FB/blog posts, TikTok carousels, YouTube Shorts, covers — so the output stops smelling like AI and comes out consistent every run. Use when a creator (not a dev) wants to set up Claude Code / an AI agent for content production across channels, or fix output that reads generic, breaks Thai fonts, or looks different every gen. Triggers: "setup AI for content", "brandbook for AI", "AI writes generic", "content creator claude md", "carousel/shorts pipeline", "cover generator".
Keep an AI character's face and voice identical across every clip. Lock the identity ONCE (train a face, lock a voice), then reuse it — never re-generate the face. Use when a recurring AI character drifts between clips, looks like a different person each render, or the brand face is never consistent.
Split AI content into two layers — AI generates raw material (backgrounds, characters, clips), code assembles the deliverable (text, layout, render). Use when AI-generated carousels/cards come out inconsistent, Thai fonts break, or every re-gen is a gamble.
Pump the brakes — clarify before acting. Use when a request is exploratory, mid-decision, or underspecified and the AI is about to edit, commit, hand off, deploy, or run a workflow on a partial picture. Fires on /hold-up or "just talk", "don't action this", "let's think first" — or auto-fires when the AI catches discuss-mode cues right as it was about to take a non-trivial action.
Game-theory review loop for AI drafts: tit-for-tat trust tracking across review rounds. Review depth adapts to the author's track record of addressing feedback, so clean drafts earn lighter reviews and ignored flags trigger deep audits. Use for recurring content or code review loops.
Personal commitments tracker with escalating confrontation. Use when the user wants accountability on recurring commitments, training, habits, or deadlines they keep deferring. Reviews the track record and escalates tone based on the avoidance pattern, not the excuses.