slack-memory-cleanup

Memory Cleanup Guide

Overview

The memory of an AI employee accumulates duplicates, misclassifications, and outdated information over time. This skill provides guides and tools to systematically organize memory.

Core principles:

No hardcoded retention period - the LLM judges based on context
Profile files (users/, channels/) follow the 1 entity = 1 file principle
If unsure, preserve rather than delete

Integration with the slack-memory-store skill:

이 skill은 slack-memory-store Organize memories stored with the skill
The folder structure, metadata schema, and type field match
Follow the principle of distinguishing profile files vs topic files

Quick Start

# 1. 분석만 (dry-run, 변경 없음)
python scripts/cleanup_memory.py {memories_path}

# 2. 결과 확인 후 실제 정리 실행
python scripts/cleanup_memory.py {memories_path} --execute

# 3. 정리 후 인덱스 업데이트
python scripts/update_index.py {memories_path}

Scripts

This skill provides scripts to assist with cleanup work.

cleanup_memory.py - Detect duplicates/misclassifications

Scan the memory folder to detect problems. By default, it only performs analysis (dry-run).

# 전체 분석
python scripts/cleanup_memory.py {memories_path}

# 특정 폴더만 분석
python scripts/cleanup_memory.py {memories_path} --folder users
python scripts/cleanup_memory.py {memories_path} --folder channels

# 상세 디버그 출력
python scripts/cleanup_memory.py {memories_path} --verbose

# 실제 정리 실행 (오분류 파일 이동)
python scripts/cleanup_memory.py {memories_path} --execute

Example output:

============================================================
📊 메모리 정리 분석 결과
============================================================

## 🔴 중복 파일

### users/ 폴더 (동일인 중복)
  email:batteryho@krafton.com:
    - 전지호 (Jiho Jeon).md (✅ 프로필)
    - 전지호 (Jiho Jeon) - 이메일 분석.md (📝 작업기록)
    - 전지호_외부플랫폼초대_2025-12-08.md (📝 작업기록)

## 🟡 오분류 파일
  전지호 (Jiho Jeon) - 이메일 분석.md
    현재: users/ → 권장: tasks/
    이유: 파일명에 작업 키워드
  Jira 티켓 조회 성공.md
    현재: channels/ → 권장: tasks/
    이유: type이 'task_completed'

## 📈 요약
  - 중복 그룹: 3개
  - 오분류 파일: 5개

Main detection features:

Distinguish profiles vs work records: In users/ duplicates, indicate which file is a profile and which is a work record
Show reason for misclassification: Provide the reason why the file was judged as misclassified
Use the type field: Check the type field in the metadata to detect mismatch with the folder

update_index.py - Update index

Update index.md after cleanup.

python scripts/update_index.py {memories_path}

Cleanup workflow

Step 1: Assess the current state

# 전체 메모리 구조 확인
ls -la {memories_path}/

# 각 폴더별 파일 수 확인
find {memories_path} -type f -name "*.md" | wc -l

# 폴더별 상세
ls -la {memories_path}/users/
ls -la {memories_path}/channels/
ls -la {memories_path}/tasks/

Step 2: Detect problems

Detect the following problems using scripts or manually:

Duplicate files - The same entity is spread across multiple files
Incorrect classification - Folder and content mismatch
Volatile information - Old files of low importance

Step 3: Execute cleanup

Appropriate actions based on detected problems:

Duplicates → Merge (seededuplication-rules.md )
Misclassification → Move (seemisclassification-rules.md )
Volatile → Delete (seecleanup-patterns.md )

Step 4: Update index

Always update the index after cleanup:

python scripts/update_index.py {memories_path}

Key cleanup targets

1. users/ folder

Normal state: 1 file per person (profile)

users/
└── 전지호 (Jiho Jeon).md    ← 프로필 파일만

Problem state: One person has multiple files

users/
├── 전지호 (Jiho Jeon).md              ← 프로필 (유지)
├── 전지호 (Jiho Jeon) - 이메일 분석.md  ← tasks/로 이동
├── 전지호 - AI 보고서.md               ← tasks/ 또는 misc/로 이동
└── Serin_Kim_김세린.md                 ← 기존 김세린 파일과 병합

Criteria:

email or user_idis the same, they are the same person
Keep only 1 profile file in users/
Move the rest to appropriate folders based on content

2. channels/ folder

Normal state: 1 file per channel (starting with the channel ID)

channels/
└── C08G76BB8JK_my-daily-scrum.md    ← 채널 프로필

Problem state: Files that are not channel information

channels/
├── C08G76BB8JK_my-daily-scrum.md      ← 유지
├── Jira 티켓 조회 성공.md              ← tasks/로 이동
└── 메일 조회 작업 성공.md              ← tasks/로 이동

Criteria:

channel_idexists and it is channel guidelines/information, keep it
Move work results, success cases, etc. to tasks/

3. tasks/ folder

Normal state: 1 file per task

tasks/
├── KIRA 프로젝트 작업 완료 - 2025-11-25.md
└── Tableau 데이터 조회 - 2025-12-09.md

Problem state: Duplicate similar content

tasks/
├── 7개_이메일_분석_2025-12-08.md     ← 삭제 (더 완전한 버전 있음)
├── 8개_이메일_분석_2025-12-08.md     ← 삭제
├── 9개_이메일_분석_2025-12-08.md     ← 삭제
└── 10개_이메일_분석_2025-12-08.md    ← 유지 (최종 버전)

Judgment criteria:

Intermediate results of the same task → keep only the final version
Similar titles on the same date → keep only the most complete one

4. Other folders

Folder	Cleanup criteria
`projects/`	Completed projects → can be moved to archive/
`decisions/`	Important, long-term retention
`meetings/`	Old items → can be deleted after summarizing
`misc/`	Top cleanup priority, delete old items
`external/news/`	Value decreases over time, can be deleted
`announcements/`	Old announcements can be deleted

Importance judgment (LLM criteria)

Without hardcoded retention periods, the LLM judges by considering the following.

Items to retain

Profile information (users/, channels/)
Decision records (decisions/)
In-progress projects (projects/)
Information related to recent interactions

Items that can be deleted

Intermediate outputs (when a final version exists)
Old everyday conversations (misc/)
News/announcements past their timeliness
Duplicate information

Considerations for judgment

Last modified date (updated metadata)
Related project status (in progress vs completed)
Relationships between files (related_to)
Importance of tags (urgent, important, etc.)

Cleanup execution examples

Example 1: Deduplicating users/

요청: "users 폴더 정리해줘"

1. 현황 파악
   - 전지호 관련 파일 6개 발견
   - 김세린 관련 파일 2개 발견

2. 분석
   - 전지호: 프로필 1개 + 작업 기록 5개
   - 김세린: 같은 사람 다른 이름 2개

3. 실행
   - 전지호 작업 기록 → tasks/로 이동
   - 김세린 파일 → 병합 후 1개만 유지

4. 결과 보고
   "users/ 정리 완료:
    - 전지호: 5개 파일 tasks/로 이동
    - 김세린: 2개 파일 1개로 병합"

Example 2: Cleaning up all memory

요청: "메모리 전체 정리해줘"

1. 현황 파악
   - 총 120개 파일
   - users/: 47개 (중복 의심)
   - tasks/: 40개 (중복 의심)
   - channels/: 10개 (오분류 의심)

2. 폴더별 분석 및 정리

3. 결과 보고
   "메모리 정리 완료:
    - 삭제: 15개 (중복/중간결과)
    - 이동: 8개 (오분류 수정)
    - 병합: 5개 (동일인 중복)
    - 현재 총: 97개 파일"

Safety guidelines

Confirm before deletion

Confirm with the user before deleting important files
decisions/, projects/ Especially careful when deleting
If unsure, move rather than delete

Backup recommended

Suggest a backup before bulk cleanup
cp -r {memories_path} {memories_path}_backup_{date}

Rollback possibility

Record the list of deleted files
Record the original locations of moved files

Reference Documents

For detailed rules, refer to the following documents:

cleanup-patterns.md - Cleanup patterns and LLM judgment criteria
deduplication-rules.md - Duplicate detection and merge rules
misclassification-rules.md - Misclassification detection and move rules