memory-optimization（内存优化）

Memory Optimization Skill

Quickly implement a comprehensive memory management system for AI agents based on Moltbook community best practices.

When to Use This Skill

• Context compression causes memory loss between sessions
• Need fast context recovery (currently 5-10 minutes, target <30 seconds)
• Want structured project tracking with clear separation of concerns
• Need automated daily memory maintenance
• Building knowledge graph for entity relationships
• Migrating from simple file-based memory to advanced system

What This Skill Provides

1. TL;DR Summary System - 30-second context recovery
2. Three-File Pattern - Structured project tracking
3. Fixed Tags System - Quick grep search capability
4. Daily Cleanup Script - 3-minute automated maintenance
5. HEARTBEAT Integration - Mandatory memory checklist
6. Rolling Summary Template - Concise daily summaries
7. Testing Framework - 6 automated tests
8. Knowledge Graph - 18 entities, 15 relationships
9. Skill Usage Tracker - Track and analyze skill usage patterns

Quick Start

TL;DR Summary System

Add to each daily log (memory/YYYY-MM-DD.md):

## ⚡ TL;DR Summary

**Core Achievements**:
- ✅ Achievement 1
- ✅ Achievement 2

**Today's Key Points**:
- Key point 1
- Key point 2

**Decisions**: Important decision made today

Three-File Pattern

For complex projects, create:

• memory/task_plan.md - What to do (goals, phases, decisions)
• memory/findings.md - What discovered (research, key info)
• memory/progress.md - What done (timeline, errors)

Fixed Tags

Use consistent tags across files:

• #memory - Memory-related content
• #decision - Important decisions
• #improvement - Optimization work
• #daily-log - Daily log entries

Daily Cleanup

Run automated cleanup:

./memory/daily-cleanup.sh

HEARTBEAT Integration

Add to HEARTBEAT.md:

### 🧠 Memory Management Checklist

Every Session Start:
- [ ] Read SOUL.md (agent identity)
- [ ] Read USER.md (user preferences)
- [ ] Read memory/YYYY-MM-DD.md (today + yesterday)
- [ ] Read MEMORY.md (long-term memory)

Scripts

See scripts/README.md for detailed usage:

• daily-cleanup.sh - 3-minute daily memory maintenance
• test-memory-system.sh - Verify all improvements working
• memory_ontology.py - Knowledge Graph management tool (with Phase 6 retrieve command)
• kg_type_fixer.py - Fix entities missing type field by inferring from ID prefix
• kg_extractor.py - KG extraction from agent sessions (LLM-driven)
• write_time_gating.py - Phase 8: Write-time gating evaluation
• archived_memory_store.py - Cold storage management (Phase 8)
• preference_engine.py - Phase 2: User preference inference from conversation history
• consolidation_engine.py - Phase 3: consolidate similar episodes into SkillCards
• decay_engine.py - Batch decay engine for memory strength management
• entity_dedup.py - Phase 2: Embedding-based entity deduplication and merging
• working_memory.py - Phase 5: Context Window layered compression (3 levels)
• memory_loader.py - Phase 6: Proactive memory recovery (3-stage staged loading)
• memory_dashboard.py - Phase 7: Memory health dashboard (Health Score, decay forecast)

Phase 6: Value-Aware Retrieval

KG entities can be retrieved by value score (weighted combination of reliability, strength, significance, preference match, recency):

# Retrieve entities sorted by value score
python3 scripts/memory_ontology.py retrieve --types Decision Finding --min-score 0.4

# Query by text and filter by score
python3 scripts/memory_ontology.py retrieve --query "architecture" --show-scores

# CLI help
python3 scripts/memory_ontology.py retrieve --help

References

See reference files for detailed guidance:

• references/implementation.md - Complete implementation guide
• references/templates.md - TL;DR, Three-file, Rolling summary templates
• references/knowledge-graph.md - KG schema and usage guide

Key Metrics

Key Insights from Moltbook

"Forget is a survival mechanism" - Compression forces distillation of experience into most resilient forms

"Knowledge graph is an index for your brain" - Query efficiency 10x better than grep

"Record immediately, not wait" - Details fade quickly

"Focus on why, not what" - Rationale is more important than the fact

File Structure

memory/
├── YYYY-MM-DD.md          # Daily log with TL;DR
├── task_plan.md            # Task planning
├── findings.md             # Research findings
├── progress.md             # Progress tracking
├── rolling-summary-template.md
├── daily-cleanup.sh
├── test-memory-system.sh
└── ontology/
    ├── memory-schema.yaml
    ├── entity-templates.md
    ├── INTEGRATION.md
    └── graph.jsonl

scripts/
└── memory_ontology.py

Skill Usage Tracker

Track and analyze skill usage patterns via Knowledge Graph:

# Record a skill usage
python3 scripts/skill_tracker.py record --skill coding-agent --status success --duration 2.5

# View usage statistics
python3 scripts/skill_tracker.py stats

# Scan session history and save to KG
python3 scripts/skill_tracker.py scan --save

Features:

• Record skill invocations with status, duration, and error info
• Automatic skill categorization (feishu/apple/coding/memory/api/system/utility)
• Session history scanning for automatic usage detection
• KG-backed storage with fallback to file

Usage Examples

Create New Daily Log with TL;DR

# Daily Memory - 2026-03-13

## ⚡ TL;DR Summary

**Core Achievements**:
- ✅ Completed task 1
- ✅ Completed task 2

**Today's Key Points**:
- Working on project X
- Found solution Y

**Decisions**: Chose approach Z

Use Knowledge Graph

# Create a decision entity
python3 scripts/memory_ontology.py create --type Decision --props '{"title":"...","rationale":"...","made_at":"...","confidence":0.9,"tags":["#decision"]}'

# Query by tags
python3 scripts/memory_ontology.py query --tags "#memory" "#decision"

# Get related entities
python3 scripts/memory_ontology.py related --id dec_xxx

Environment Variables

# GLM API configuration (used by kg_extractor.py)
export OPENAI_API_KEY="your-glm-token"
export OPENAI_BASE_URL="https://open.bigmodel.cn/api/paas/v4"
export OPENAI_MODEL="glm-5"

# Global KG path (optional, defaults to ~/.openclaw/workspace/memory/ontology)
# Configure via KG_DIR environment variable

KG Sharing Across Agents

Multi-agent setups can share a single Knowledge Graph for collaborative memory.

Setup

# 1. Create shared directory
mkdir -p ~/.openclaw/shared-kg

# 2. Create symlink to main KG
ln -sf ~/.openclaw/workspace/memory/ontology/graph.jsonl ~/.openclaw/shared-kg/main-kg.jsonl

Usage by Agents

Each agent should reference the shared KG in their TOOLS.md:

## Knowledge Graph (KG)

- **共享 KG**: ~/.openclaw/shared-kg/main-kg.jsonl
- 包含所有重要的长期记忆、决策、经验教训

Script Usage with Shared KG

# Direct KG path
KG_DIR=~/.openclaw/shared-kg python3 scripts/memory_ontology.py query --tags "#decision"

# Or set in .env
KG_DIR=~/.openclaw/shared-kg/

Benefits:

• All agents access the same entity pool
• Decisions and lessons are shared across agents
• No duplicate entity creation

OpenClaw Skill Invocation

When the user types /xmo, OpenClaw automatically invokes this memory-optimization skill.

Add the following to OpenClaw's settings.json or skills.json:

{
  "skills": {
    "xmo": {
      "path": "./memory-optimization",
      "description": "Memory optimization skill for AI agents"
    }
  }
}

After configuration, the user can activate this skill by typing /xmo.

Next Steps

1. Run test script: ./memory/test-memory-system.sh
2. Verify TL;DR exists in today's log
3. Start using KG for important decisions
4. Run daily cleanup each day

For complete implementation details, see references/implementation.md.