记忆索引管理器

记忆索引系统。用于管理长期项目记忆、自动维护索引、快速定位历史项目。仅在用户明确说"回忆一下""remember""recall"时触发读取，写入与OpenClaw原生机制保持一致，索引维护在后台自动进行。

Memory Index Manager

A skill for project continuity and memory reconstruction. Maintains a curated index of topics with multi-generation consolidation to help recover context across long-running conversations.

Core Design

Trigger Rules

Operation	Trigger	Behavior
Read	User explicitly says "回忆一下" / "remember" / "recall" / "记不记得"	Check INDEX.md → read consolidation → respond
Write	OpenClaw native flush (pre-compaction)	Write to memory/YYYY-MM-DD.md as usual
Index Update	End of conversation + heartbeat	Propose new entries, update dates, consolidate if needed
Daily Flush	Every day at 11:00 AM	Archive yesterday's memory, update INDEX, check consolidation

Philosophy

• Passive by default: Doesn't interrupt normal workflow
• Active on request: Full context recovery when explicitly asked
• Automatic maintenance: Index upkeep happens in background

---

File Structure

text

memory/
├── INDEX.md                    # Master index of tracked topics
├── MEMORY.md                   # Native long-term memory (OpenClaw)
├── YYYY-MM-DD.md               # Daily memories (OpenClaw native writes)
└── consolidated/               # Multi-generation consolidations
    ├── {topic}-gen1-YYYY-MM-DD.md
    ├── {topic}-gen2-YYYY-MM-DD.md
    └── ...

---

INDEX.md Format

markdown

# Memory Index

## Projects
- **project-slug**
  - Dates: [2025-03-20, 2025-03-22, 2025-03-25, 2025-04-01]
  - Consolidations:
    - gen1 (2025-03-25): consolidated/project-slug-gen1-2025-03-25.md
  - Current: consolidated/project-slug-gen1-2025-03-25.md
  - Status: active | completed | paused | abandoned

## Major Events
- **event-slug**: [2025-03-15]
  - Description: Brief description

## Decisions
- **decision-slug**: [2025-03-10, 2025-03-12]
  - Consolidations:
    - gen1 (2025-03-12): consolidated/decision-slug-gen1-2025-03-12.md
  - Current: consolidated/decision-slug-gen1-2025-03-12.md

---

Workflows

1. Explicit Recall ("回忆一下")

User trigger: "回忆一下那个网站项目" / "Remember the website project?"

Agent workflow:

text

1. Read INDEX.md
2. Find matching topic (fuzzy match on name)
3. If found:
   a. Read Current consolidation file
   b. Read today's memory/YYYY-MM-DD.md for fresh updates
   c. Synthesize and respond
4. If not found in INDEX:
   a. Fall back to memory_search("website project")
   b. Respond with search results
   c. (Optional) Propose adding to index if significant

2. Native Write (Automatic)

Trigger: OpenClaw pre-compaction flush or user says "remember this"

Behavior: Write to memory/YYYY-MM-DD.md per OpenClaw native rules

No skill involvement - this is pure OpenClaw behavior.

3. Index Maintenance (Background)

Trigger: End of conversation OR heartbeat

Agent workflow:

text

1. Review today's memory file
2. Identify topics worth indexing:
   - New projects started
   - Major events occurred
   - Important decisions made
   - Existing indexed topics mentioned
3. For existing topics: add today's date to Dates list
4. Propose new index entries to user (if autoPropose enabled)
5. Check consolidation threshold:
   - Count dates since last consolidation
   - If > 3: propose consolidation to user
   - On approval: create gen{N+1} consolidation

4. Daily Flush (11:00 AM)

Trigger: System cron job at 11:00 AM daily

Agent workflow:

text

1. Identify yesterday's date (YYYY-MM-DD)
2. Check if yesterday's memory file exists
   - If not: create empty file with <!-- DAILY_FLUSH --> marker
3. Execute OpenClaw native memory storage:
   - Append <!-- NATIVE_FLUSH --> to yesterday's file
   - Archive any in-memory context from yesterday
4. Execute Skill index update:
   - Scan yesterday's file for index-worthy topics
   - Update INDEX.md with yesterday's date
   - Check if any topic needs consolidation (>3 days)
5. Initialize today's memory file (if not exists)
6. Log flush completion

Daily Flush Script: ~/.openclaw/scripts/daily-flush.sh Launchd Config: ~/Library/LaunchAgents/com.openclaw.daily-flush.plist

Purpose: Ensure yesterday's memory is properly archived even if:

• User didn't explicitly say "remember"
• Session didn't reach memoryFlush threshold
• System was restarted

---

Consolidation Rules

When to Consolidate

A topic needs new consolidation when:

• It has >3 dates (excluding today) since last consolidation

Consolidation Process

1. Create new file: consolidated/{topic}-gen{N}-{today}.md
2. Sources:
   • Previous generation consolidation (if N>1)
   • Daily memories from dates since last consolidation
3. Content structure:

   markdown

   # {Topic} - Generation {N} Consolidation
   
   **Created**: 2025-04-15
   **Covers**: 2025-04-01 to 2025-04-15
   **Previous**: consolidated/{topic}-gen{N-1}-{date}.md
   
   ## Summary
   [Compiled narrative of all covered periods]
   
   ## Key Points
   - Point 1
   - Point 2
   
   ## Decisions Made
   - Decision 1
   
   ## Current Status
   [Where things stand now]
   
   ## Source References
   - gen{N-1}: [key info from previous consolidation]
   - 2025-04-01: [what happened]
   - 2025-04-05: [what happened]
   - ...
   

4. Mark sources archived: Add <!-- ARCHIVED: consolidated into {file} --> to daily files
5. Update INDEX.md: Add consolidation entry, update Current pointer

---

Categories

Category	Description	Example
Projects	Multi-day efforts with deliverables	website redesign, app development
Major Events	Significant life/career changes	job change, relocation
Decisions	Choices with lasting impact	tech stack, vendor selection
Preferences	Personal settings that persist	communication style, tools

---

Configuration

memory-index-manager/config.json:

json

{
  "consolidationThreshold": 3,
  "autoProposeIndex": true,
  "categories": ["Projects", "Major Events", "Decisions", "Preferences"],
  "indexFile": "memory/INDEX.md",
  "consolidatedDir": "memory/consolidated",
  "recallTriggers": ["回忆一下", "remember", "recall", "记不记得"]
}

---

Integration with OpenClaw Native Memory

Aspect	Native OpenClaw	This Skill
Storage	memory/*.md	INDEX.md + consolidated/*.md
Write timing	Auto (flush) + Manual	Same (no change)
Read timing	Any search	Only on explicit recall trigger
Search method	Semantic (memory_search)	Index lookup + targeted read
Granularity	Chunk-level (~400 tokens)	Topic-level consolidation

Key point: This skill's files are stored in memory/ and indexed by OpenClaw's semantic search, but the skill only activates on explicit recall requests.

---

Example Session

text

[Day 1 - 2025-03-20]
User: "Let's start a new project: building a personal blog"
Agent: [writes to memory/2025-03-20.md]
Agent: "I'll track this as 'personal-blog' project. Adding to index?"
User: "Yes"
→ Creates INDEX.md entry: personal-blog, Dates: [2025-03-20]

[Day 2 - 2025-03-22]
User: "For the blog, let's use Next.js"
Agent: [writes to memory/2025-03-22.md]
→ Updates INDEX.md: personal-blog Dates: [2025-03-20, 2025-03-22]

[Day 3 - 2025-03-25]
User: "Blog hosting should be Vercel"
Agent: [writes to memory/2025-03-25.md]
→ Updates INDEX.md: personal-blog Dates: [2025-03-20, 2025-03-22, 2025-03-25]
→ TRIGGER: 3 dates, proposes consolidation
User: "Go ahead"
→ Creates consolidated/personal-blog-gen1-2025-03-25.md
→ Archives daily files

[Day 7 - 2025-04-05]
User: "回忆一下博客项目"
Agent: [reads INDEX.md → finds personal-blog]
Agent: [reads consolidated/personal-blog-gen1-2025-03-25.md]
Agent: "We started the personal blog project on March 20th. We decided to use Next.js 
        (March 22nd) and host on Vercel (March 25th). The project is currently active."

---

Migration from v1.0.0

Breaking Changes:

• 旧版本触发条件（每天9:00 AM、创建项目时等）已移除
• 改为纯按需触发（"回忆一下"），减少干扰
• 索引格式兼容，可无缝迁移

升级步骤：

1. 备份现有的 memory/INDEX.md
2. 更新 skill 到 2.0.0
3. 继续使用，索引文件自动兼容

---

Usage Examples

User Says	Agent Action
"回忆一下那个网站项目"	查 INDEX.md → 读 consolidation → 回答
"remember the blog project"	同上，支持英文触发
"把这个加入索引"	询问分类 → 创建 INDEX.md 条目
"显示索引"	展示 INDEX.md 内容
"Consolidate blog project"	检查阈值 → 执行整合

---

Notes

• Index entries are append-only; mark deprecated rather than delete
• Consolidated files are immutable; new generations for updates
• Daily memories marked archived are preserved but skipped by default
• User can force deep read: "check the original March 20th file"