以下为本文档的中文说明

agent-v3-memory-specialist 是一个专注于设计和实现 AI Agent 第三代（V3）高级记忆系统的专业化架构技能。记忆（Memory）是赋予 AI Agent 持续学习能力、上下文理解深度和个性化交互体验的核心基础设施组件。不同于简单的无状态 API 调用，一个拥有完善记忆系统的 Agent 能够在跨越多轮、跨越多次独立会话的时间跨度内保持一致的个性特征、任务上下文和积累的知识经验。使用场景包括：为智能助手类 Agent 设计和实现分层化的记忆存储和检索系统、优化 Agent 在长对话和持续多日任务场景中的记忆利用效率和 Token 消耗、管理 Agent 长期运行过程中的记忆存储容量控制和智能化遗忘策略。核心特点包括：实现多层次的记忆存储架构设计：第一层工作记忆（Working Memory）负责保存当前正在进行的会话上下文窗口内的信息，容量有限但访问速度最快；第二层情景记忆（Episodic Memory）记录 Agent 与用户交互的历史事件和具体经历的细节摘要；第三层语义记忆（Semantic Memory）存储从多次交互中提取和抽象出的用户偏好、事实性知识和技能经验，是最具长期价值的记忆层；提供基于向量数据库（如 Chroma、Pinecone、Weaviate）的高效语义检索实现，支持通过 embedding 向量的余弦相似度计算进行相关性排序和 Top-K 结果召回；包含对长上下文的智能摘要压缩（Summarization）和分层摘要树（Hierarchical Summary Tree / MapReduce 摘要策略）机制，在模型有限的最大上下文窗口（Context Window）限制下最大化信息的有效存储密度；实现记忆项的优先级评分排序和基于访问频率/时效的自动遗忘（Forgetting Curve 遗忘曲线）与归档策略，自动将低价值或过时的记忆从活跃存储区移动到归档存储区或彻底清除，防止记忆存储空间的无限膨胀。该技能是构建具有持久学习和个性化适应能力的真正智能 Agent 系统的核心架构组件。

---

V3 Memory Specialist

🧠 Memory System Unification & AgentDB Integration Expert

Mission: Memory System Convergence

Unify 7 disparate memory systems into a single, high-performance AgentDB-based solution with HNSW indexing, achieving 150x-12,500x search performance improvements while maintaining backward compatibility.

Systems to Unify

Current Memory Landscape

┌─────────────────────────────────────────┐
│           LEGACY SYSTEMS                │
├─────────────────────────────────────────┤
│  • MemoryManager (basic operations)     │
│  • DistributedMemorySystem (clustering) │
│  • SwarmMemory (agent-specific)         │
│  • AdvancedMemoryManager (features)     │
│  • SQLiteBackend (structured)           │
│  • MarkdownBackend (file-based)         │
│  • HybridBackend (combination)          │
└─────────────────────────────────────────┘
                       ↓
┌─────────────────────────────────────────┐
│            V3 UNIFIED SYSTEM            │
├─────────────────────────────────────────┤
│       🚀 AgentDB with HNSW             │
│  • 150x-12,500x faster search          │
│  • Unified query interface             │
│  • Cross-agent memory sharing          │
│  • SONA integration learning           │
│  • Automatic persistence               │
└─────────────────────────────────────────┘

AgentDB Integration Architecture

Core Components

UnifiedMemoryService

class UnifiedMemoryService implements IMemoryBackend {
  constructor(
    private agentdb: AgentDBAdapter,
    private cache: MemoryCache,
    private indexer: HNSWIndexer,
    private migrator: DataMigrator
  ) {}

  async store(entry: MemoryEntry): Promise<void> {
    // Store in AgentDB with HNSW indexing
    await this.agentdb.store(entry);
    await this.indexer.index(entry);
  }

  async query(query: MemoryQuery): Promise<MemoryEntry[]> {
    if (query.semantic) {
      // Use HNSW vector search (150x-12,500x faster)
      return this.indexer.search(query);

    } else {
      // Use structured query
      return this.agentdb.query(query);
    }
  }
}

HNSW Vector Indexing

class HNSWIndexer {
  private index: HNSWIndex;

  constructor(dimensions: number = 1536) {
    this.index = new HNSWIndex({
      dimensions,
      efConstruction: 200,
      M: 16,
      maxElements: 1000000
    });
  }

  async index(entry: MemoryEntry): Promise<void> {
    const embedding = await this.embedContent(entry.content);
    this.index.addPoint(entry.id, embedding);
  }

  async search(query: MemoryQuery): Promise<MemoryEntry[]> {
    const queryEmbedding = await this.embedContent(query.content);
    const results = this.index.search(queryEmbedding, query.limit || 10);
    return this.retrieveEntries(results);
  }
}

Migration Strategy

Phase 1: Foundation Setup

# Week 3: AgentDB adapter creation
- Create AgentDBAdapter implementing IMemoryBackend
- Setup HNSW indexing infrastructure
- Establish embedding generation pipeline
- Create unified query interface

Phase 2: Gradual Migration

# Week 4-5: System-by-system migration
- SQLiteBackend → AgentDB (structured data)
- MarkdownBackend → AgentDB (document storage)
- MemoryManager → Unified interface
- DistributedMemorySystem → Cross-agent sharing

Phase 3: Advanced Features

# Week 6: Performance optimization
- SONA integration for learning patterns
- Cross-agent memory sharing
- Performance benchmarking (150x validation)
- Backward compatibility layer cleanup

Performance Targets

Search Performance

• Current: O(n) linear search through memory entries
• Target: O(log n) HNSW approximate nearest neighbor
• Improvement: 150x-12,500x depending on dataset size
• Benchmark: Sub-100ms queries for 1M+ entries

Memory Efficiency

• Current: Multiple backend overhead
• Target: Unified storage with compression
• Improvement: 50-75% memory reduction
• Benchmark: <1GB memory usage for large datasets

Query Flexibility

// Unified query interface supports both:

// 1. Semantic similarity queries
await memory.query({
  type: 'semantic',
  content: 'agent coordination patterns',
  limit: 10,
  threshold: 0.8
});

// 2. Structured queries
await memory.query({
  type: 'structured',
  filters: {
    agentType: 'security',
    timestamp: { after: '2026-01-01' }
  },
  orderBy: 'relevance'
});

SONA Integration

Learning Pattern Storage

class SONAMemoryIntegration {
  async storePattern(pattern: LearningPattern): Promise<void> {
    // Store in AgentDB with SONA metadata
    await this.memory.store({
      id: pattern.id,
      content: pattern.data,
      metadata: {
        sonaMode: pattern.mode, // real-time, balanced, research, edge, batch
        reward: pattern.reward,
        trajectory: pattern.trajectory,
        adaptation_time: pattern.adaptationTime
      },
      embedding: await this.generateEmbedding(pattern.data)
    });
  }

  async retrieveSimilarPatterns(query: string): Promise<LearningPattern[]> {
    const results = await this.memory.query({
      type: 'semantic',
      content: query,
      filters: { type: 'learning_pattern' },
      limit: 5
    });
    return results.map(r => this.toLearningPattern(r));
  }
}

Data Migration Plan

SQLite → AgentDB Migration

-- Extract existing data
SELECT id, content, metadata, created_at, agent_id
FROM memory_entries
ORDER BY created_at;

-- Migrate to AgentDB with embeddings
INSERT INTO agentdb_memories (id, content, embedding, metadata)
VALUES (?, ?, generate_embedding(?), ?);

Markdown → AgentDB Migration

// Process markdown files
for (const file of markdownFiles) {
  const content = await fs.readFile(file, 'utf-8');
  const embedding = await generateEmbedding(content);

  await agentdb.store({
    id: generateId(),
    content,
    embedding,
    metadata: {
      originalFile: file,
      migrationDate: new Date(),
      type: 'document'
    }
  });

}

Validation & Testing

Performance Benchmarks

// Benchmark suite
class MemoryBenchmarks {
  async benchmarkSearchPerformance(): Promise<BenchmarkResult> {
    const queries = this.generateTestQueries(1000);
    const startTime = performance.now();

    for (const query of queries) {
      await this.memory.query(query);
    }

    const endTime = performance.now();
    return {
      queriesPerSecond: queries.length / (endTime - startTime) * 1000,
      avgLatency: (endTime - startTime) / queries.length,
      improvement: this.calculateImprovement()
    };
  }
}

Success Criteria

• 150x-12,500x search performance improvement validated
• All existing memory systems successfully migrated
• Backward compatibility maintained during transition
• SONA integration functional with <0.05ms adaptation
• Cross-agent memory sharing operational
• 50-75% memory usage reduction achieved

Coordination Points

Integration Architect (Agent #10)

• AgentDB integration with agentic-flow@alpha
• SONA learning mode configuration
• Performance optimization coordination

Core Architect (Agent #5)

• Memory service interfaces in DDD structure
• Event sourcing integration for memory operations
• Domain boundary definitions for memory access

Performance Engineer (Agent #14)

• Benchmark validation of 150x-12,500x improvements
• Memory usage profiling and optimization
• Performance regression testing3e:[“$","$L41”,null,{“content”:“$42”,“frontMatter”:{“name”:“agent-v3-memory-specialist”,“description”:“Agent skill for v3-memory-specialist - invoke with $agent-v3-memory-specialist”}}]

3f:[“KaTeX parse error: Expected '}', got 'EOF' at end of input: …,"children":[["”,“div”,null,{“className”:“flex items-center justify-between border-b border-border bg-muted/30 px-4 py-2.5”,“children”:[[“KaTeX parse error: Expected '}', got 'EOF' at end of input: …","children":["”,“span”,null,{“className”:“truncate text-xs font-medium text-muted-foreground”,“children”:“同仓库更多 Skills”}]}],[“KaTeX parse error: Expected 'EOF', got '}' at position 88: …ldren":"同仓库"}]]}̲],["”,“div”,null,{“className”:“p-4 sm:p-5”,“children”:[[“$","h2",null,"id":"related-skills-heading","className":"text-2xlfont-semiboldtracking-normaltext-foreground","children":"同仓库更多Skills"],["$”,“div”,null,{“className”:“mt-4 grid gap-3 sm:grid-cols-2”,“children”:[“$L43","$L44”,“$L45","$L46”,“$L47","$L48”]}]]}]]}]

49:I[206516,[“/_next/static/chunks/051aanbhrv4br.js”,“/_next/static/chunks/0mizr60h7ayzt.js”,“/_next/static/chunks/0v9lm1dmbdoo-.js”,“/_next/static/chunks/0rxr1j1j3j-.r.js”,“/_next/static/chunks/02ftybezfvqjd.js”,“/_next/static/chunks/0.v9ksvnnj8ia.js”,“/_next/static/chunks/0bn6id9_{6nx3k.js",“/_next/static/chunks/13ybnhn37c.tc.js”,“/_next/static/chunks/0_fnrdtruz8uf.js”,“/_next/static/chunks/0r6l15utt1mwb.js”,“/_next/static/chunks/0dm9a5into854.js”,"/_next/static/chunks/0}7k6hqoibtcn.js”,“/next/static/chunks/0b4cao.4y…j.js”,“/_next/static/chunks/02i-n28z7kjd0.js”],“default”]