somnus0x/agt-skill-pack/ai-memory

AI Memory — Give Your AI a Past and a Present

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:

1. Long-term memory — search past sessions by keyword (FTS5 index)
2. Short-term memory — know current project state at session start (hot cache)

The Problem (two sides)

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.

Setup: Long-Term Memory (Session Search)

Drop-in prompt (copy into Claude Code):

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.

Usage

# 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

Manual setup (if you prefer)

Key components:

1. Parser — Read JSONL session files, extract type: "human" and type: "assistant" messages
2. SQLite + FTS5 — Virtual table with tokenize='unicode61' for multilingual support
3. CLI — search "keyword", ingest (rebuild index), stats (show counts)

No cloud, no API keys, no vector DB, no embeddings. FTS5 is built into Python's sqlite3.

Setup: Short-Term Memory (Hot Cache)

Drop-in prompt (copy into Claude Code):

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.

What hot-cache.md looks like

# 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

Keeping it fresh

• Manual: Before closing a session, tell the AI: "update hot-cache.md with what we did today."
• Automated: Set up a cron to refresh periodically:

0 */4 * * * cd /path/to/project && claude -p "Read [source files]. Update hot-cache.md." >> /dev/null 2>&1

Wire It Up — Add to CLAUDE.md

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.

When to use

• "What did we decide about X?" — search past sessions
• "I wrote a good prompt for Y" — search and recover it
• "What shipped yesterday?" — hot cache knows
• Starting a new session on a multi-day task — hot cache has continuity
• Before making architectural decisions — search for past discussions on the topic

Technical notes

• Everything local. No cloud, no API keys, no external services.
• No vector DB. FTS5 full-text search handles keyword lookup without embeddings.
• Multilingual. FTS5 unicode61 tokenizer works with Thai, English, and mixed text.
• Fast. 800+ sessions (3M+ chars) searched in milliseconds.
• Universal. Works with Claude Code, Cursor, Copilot, Windsurf, Aider — any tool that reads files.