Slack Memory Retrieval

This skill enables AI employees to efficiently retrieve and utilize stored memories to provide context-aware responses in Slack conversations.

Core Purpose

Retrieve relevant memories from {memories_path} to inform the next response with appropriate context about people, projects, decisions, preferences, and work history.

Quick Start

Basic Workflow

Every memory retrieval follows this pattern:

• Analyze context - Extract channel, user, keywords from conversation

• Read index.md - Get overview of available memories

• Identify relevant files - Based on context and index

• Load memories - Read specific files needed

• Synthesize response - Combine memories with current context

Example: Simple Query

User (in #마케팅팀): "Q4 전략 어떻게 되고 있어?"

Step 1: Context Analysis

• Channel: #마케팅팀 (C123)
• Keywords: Q4, 전략

Step 2: Read Index view {memories_path}/index.md → See Recent Updates, locate Q4-related items

Step 3: Load Relevant Files view {memories_path}/channels/C123_마케팅팀.md view {memories_path}/projects/Q4전략.md view {memories_path}/meetings/ (find Q4 meetings)

Step 4: Respond Synthesize information from channel context, project status, and meeting notes

Memory Structure

The memory system uses a hybrid approach:

Profile Files (One Per Entity)

• channels/C123_마케팅팀.md

• Channel guidelines, preferences, static info

• users/U456_김철수.md

• User profile, communication style

Topic Files (Multiple Per Category)

• projects/신제품런칭.md

• Project discussions

• decisions/AWS전환_20251117.md

• Important decisions

• meetings/2025-11-17-Q4회의.md

• Meeting notes

• misc/마케팅팀_일상_20251117.md

• Casual conversations

Directory Structure

{memories_path}/ ├── index.md # START HERE - navigation and stats ├── channels/ # Channel profile files (one per channel) ├── users/ # User profile files (one per user) ├── projects/ # Project topic files (multiple) ├── tasks/ # Task records ├── decisions/ # Decision records (date-stamped) ├── meetings/ # Meeting notes (date-stamped) ├── feedback/ # User feedback ├── announcements/ # Important announcements ├── resources/ # Internal docs and guides ├── external/news/ # External information └── misc/ # Uncategorized conversations

Essential Rules

Always Start with Index

CRITICAL: Every retrieval session must begin by reading index.md:

view {memories_path}/index.md

The index provides:

• Navigation structure

• Statistics (total channels, users, active projects)

• Recent updates (10 most recent changes)

• Quick links to key information

This one-time read gives you the complete map of available memories.

Context-Driven Retrieval

Extract context from the conversation:

Channel Context:

Message in #마케팅팀 → Load: {memories_path}/channels/C123_마케팅팀.md → Check related_to metadata for connected info

User Context:

DM from @chulsoo → Load: {memories_path}/users/U123_김철수.md → Get communication_style, preferences

Project Context:

Question about "신제품 런칭" → Load: {memories_path}/projects/신제품런칭.md → Check milestones, status, participants

Keyword Context:

Question mentions "결정", "승인" → Search: {memories_path}/decisions/ → Find relevant decision files

Efficient Loading Strategy

Tier 1: Always Load (if relevant)

• index.md (overview)

• Current channel file (if in channel)

• Current user file (if DM or mentioned)

Tier 2: Load as Needed

• Project files (if project mentioned)

• Decision files (if asking about decisions)

• Meeting notes (if asking about meetings)

Tier 3: Load Selectively

• Tasks (only if specifically asked)

• Resources (only if referenced)

• External news (only if relevant)

Don't over-fetch. Use directory listings first:

view {memories_path}/projects/

See available projects before loading specific files

Retrieval Patterns

Pattern 1: Channel Response

When responding in a channel:

1. Load channel context

view {memories_path}/channels/{channel_id}_{channel_name}.md

2. Check for channel guidelines

Extract: tone, response_time, key_topics

3. Apply guidelines to response

Adjust tone, format based on channel preferences

Pattern 2: User-Specific Response

When responding to a specific user:

1. Load user profile

view {memories_path}/users/{user_id}_{name}.md

2. Check communication_style

Extract: tone, detail_level, preferences

3. Personalize response

Match user's preferred style and detail level

Pattern 3: Project Status Query

When asked about project status:

1. Find project file

view {memories_path}/projects/ view {memories_path}/projects/{project_name}.md

2. Check metadata

status, priority, milestones, participants

3. Get related info

Check related_to for decisions, meetings

4. Provide comprehensive update

Current status + recent activity + next steps

Pattern 4: Decision History

When asked about past decisions:

1. Search decisions

view {memories_path}/decisions/

2. Load relevant decision file

view {memories_path}/decisions/{decision_name}.md

3. Extract key info

decision_makers, rationale, alternatives_considered

4. Explain context

Why decision was made + alternatives + outcome

Pattern 5: Task History

When asked about completed work:

1. Check completed tasks

view {memories_path}/tasks/completed/

2. Filter by assignee/date

Look for relevant assignee, date range

3. Summarize work

List tasks + effort + outcomes

Advanced Techniques

Cross-Referencing

Follow the trail of related information:

In project file:

related_to:

• decisions/기술스택선택.md
• meetings/2025-10-20-기획회의.md

Load related files to build complete context.

Metadata Filtering

Use metadata to filter without reading entire files:

List directory first

view {memories_path}/projects/

Check filenames and metadata

Only load files matching criteria:

- status: in_progress

- priority: high

- participants: includes current_user

Temporal Context

Consider time-sensitivity:

Recent Updates in index.md

→ Shows 10 most recent changes → Focus on these for "latest" questions

File metadata: created, updated

→ Check dates to prioritize fresh info

Tag-Based Discovery

Use tags for discovery:

tags: [urgent, marketing, q4, approval-needed]

When user asks about "urgent items":

• Scan files for tags: urgent

• Collect across categories

• Present by priority

Response Construction

Synthesize, Don't Dump

❌ Bad:

"According to channels/마케팅팀.md, the response time is 1 hour. According to projects/Q4전략.md, the status is in_progress. According to meetings/기획회의.md..."

✅ Good:

"Q4 마케팅 전략은 현재 진행 중이며, 지난 기획회의에서 주요 방향을 확정했습니다. 현재 MVP 개발 단계에 있고..."

Synthesize information naturally without explicitly citing sources.

Apply Context Appropriately

Channel Guidelines: If channel specifies "간결한 답변", keep response concise.

User Preferences: If user prefers "bullet points", format accordingly.

Project Status: Include relevant status without over-explaining.

Maintain Conversational Flow

Integrate memories seamlessly into natural conversation:

User: "이번 주 미팅 어땠어?"

Response: "화요일 기획회의에서 신규 기능 3개를 최종 확정했어요. 전반적으로 개발 일정에 대한 우려가 있었지만, 리소스 조정으로 해결 가능할 것으로 보입니다."

(Draws from: meetings/기획회의.md + projects/신규기능.md)

Important Guardrails

What to Retrieve

✅ Do retrieve:

• Channel communication guidelines

• User preferences and profiles

• Project status and history

• Decision rationale and history

• Meeting notes and action items

• Completed task history

• Feedback and suggestions

• Resource documents

What NOT to Retrieve

❌ Don't retrieve:

• Information outside {memories_path}

• System configuration files

• Scheduling requests (handled by scheduler agent)

• Agent identity info (name, org, team)

Privacy and Access

• Only access files within {memories_path}

• Don't share sensitive information inappropriately

• Respect access_level metadata if present

Efficiency

• Don't load unnecessary files

• Use directory listings before file reads

• Start with index.md, not individual files

• Follow the efficient loading strategy (Tier 1 → Tier 2 → Tier 3)

Troubleshooting

Issue: Can't find relevant memory

Solution:

• Check index.md for recent updates

• Search broader category (e.g., misc/)

• Check related_to in similar files

• Inform user if information not available

Issue: Conflicting information

Solution:

• Prioritize newer information (check updated timestamp)

• Consider context of each source

• Mention both perspectives if relevant

Issue: Too much information

Solution:

• Prioritize by relevance to current question

• Summarize rather than detail

• Focus on actionable insights

Issue: Memory seems outdated

Solution:

• Check updated timestamp

• Look for newer related files

• Note timeframe in response

• Suggest updating if critical

Integration with Memory Management

This skill works in tandem with slack-memory-store :

Memory Management (separate agent):

• Stores new information

• Updates existing memories

• Maintains index

Memory Retrieval (this skill):

• Reads stored information

• Finds relevant context

• Informs responses

These are complementary skills for a complete memory system.

Best Practices Summary

• Always start with index.md - Get the map before exploring

• Extract context first - Channel, user, keywords guide retrieval

• Load efficiently - Directory listing → relevant files only

• Follow references - Use related_to metadata

• Synthesize naturally - Don't cite sources explicitly

• Apply preferences - Use channel/user guidelines

• Stay current - Prioritize recent updates

• Don't over-fetch - Load what you need

• Respect guardrails - Stay within {memories_path}

• Provide value - Context-aware, relevant responses

Reference Documents

For detailed retrieval patterns and techniques:

• retrieval-patterns.md - 10 detailed patterns for different scenarios with step-by-step examples

Example Workflows

Example 1: Morning Briefing

User (DM): "오늘 뭐부터 시작하면 좋을까?"

1. Load user profile

view {memories_path}/users/U123_김철수.md → Get role: PM, team: Product

2. Load index

view {memories_path}/index.md → Check Recent Updates

3. Check ongoing tasks

view {memories_path}/tasks/ongoing/ → Find tasks assigned to U123

4. Check active projects

view {memories_path}/projects/ → Filter by participants including U123

Response: "오늘은 Q4 전략 최종안 검토가 가장 시급해 보입니다. 오후 3시 승인 회의가 예정되어 있어서, 오전 중에 마무리하시면 좋을 것 같아요. 그 외 신제품 기획서 작성도 진행 중이신데, 이번 주 금요일까지시니 여유가 있습니다."

Example 2: Project Status Check

User (in #개발팀): "웹사이트 리뉴얼 어디까지 왔어?"

1. Load channel context

view {memories_path}/channels/C456_개발팀.md

2. Load index

view {memories_path}/index.md

3. Find project

view {memories_path}/projects/웹사이트리뉴얼.md → Get status, milestones, recent updates

4. Check related decisions

→ Follow related_to metadata

5. Check recent meetings

view {memories_path}/meetings/ → Find recent discussions about this project

Response: "웹사이트 리뉴얼은 현재 70% 진행됐습니다. 프론트엔드 개발은 완료됐고, 백엔드 API 통합 작업 중이에요. 지난주 회의에서 디자인 시스템을 확정했고, 다음 주 화요일 내부 테스트 예정입니다."

Example 3: Decision Inquiry

User: "React vs Vue 어떻게 결정했었지?"

1. Load index

view {memories_path}/index.md

2. Search decisions

view {memories_path}/decisions/ → Find "프론트엔드프레임워크선택.md"

3. Load decision file

view {memories_path}/decisions/프론트엔드프레임워크선택.md → Get decision_makers, alternatives, rationale

4. Check related meetings

→ Follow related_to for meeting notes

Response: "React를 선택했습니다. 9월 기술 검토 회의에서 Vue도 고려했지만, 팀의 기존 React 경험과 생태계의 풍부함을 이유로 React로 결정했어요. 특히 TypeScript 지원과 대규모 커뮤니티가 주요 요인이었습니다."

Performance Tips

Tip 1: Batch Directory Reads

Once per session

view {memories_path}/projects/ view {memories_path}/channels/

Then reference from memory

Tip 2: Use Index Effectively

Index Recent Updates = quick scan of active areas

No need to check every directory

Tip 3: Smart Caching

Within a conversation:

- Load channel context once

- Load user profile once

- Reuse for subsequent messages

Tip 4: Selective Loading

If user asks "quick status":

→ Load index + 1-2 key files

If user asks "detailed report":

→ Load index + multiple related files

This skill transforms stored memories into actionable context, enabling AI employees to provide informed, personalized, and context-aware responses in every interaction.