Telegram AI代理记忆层

1. Connect Telegram    →  Authorize via @AgentMemoryBot
2. Add sources         →  Channels, groups, folders, topics
3. System indexes      →  History download → entity extraction → graph building
4. Get API key         →  Create in bot, use in your agent
5. Agent queries       →  search / digest / decisions / context — all via API

graph TB
    subgraph Sources["📱 Telegram Sources"]
        CH[Channels]
        GR[Groups & Topics]
        FL[Folders]
    end

    subgraph Ingestion["⚙️ Ingestion Pipeline"]
        COL[Telethon Collector] --> NF[Noise Filter]
        NF --> ME[Metadata & Threading]
        ME --> EE[Entity Extraction]
        EE --> EMB[BGE-M3 Embedding]
    end

    subgraph Memory["🧠 Memory Engine"]
        PG["PostgreSQL + ParadeDB\n📝 BM25 Full-Text"]
        MV["Milvus 2.5\n🧬 Dense + Sparse Vectors"]
        FK["FalkorDB\n🕸️ Knowledge Graph"]
    end

    subgraph Interface["🔌 Agent Interface"]
        MCP["MCP Server\n(Streamable HTTP)"]
        REST["REST API\n(FastAPI)"]
        PKG["pip package\n(agent-memory-mcp)"]
    end

    subgraph Agents["🤖 Your Agents"]
        CL[Claude Desktop / Cursor]
        CUSTOM[Custom Bots & Agents]
        ANY[Any MCP Client]
    end

    Sources --> COL
    EMB --> PG & MV & FK
    Memory --> MCP & REST
    PKG -.->|thin client| REST
    MCP --> CL
    MCP --> ANY
    REST --> CUSTOM

    subgraph TON["💎 TON Payments"]
        CR[Pay-per-query Points]
    end

    TON -.-> Interface

Query arrives → Self-RAG gate → Route decision:

├─ ⚡ Overview     Pre-computed summary exists → instant answer
│
├─ 📊 Cascaded    BM25 finds 30+ posts → entity filter → map-reduce (up to 500 posts)
│
└─ 🔍 Standard    Parallel retrieve (BM25 + Vector + Graph + Hashtag)
                      → Merge & dedup → Rerank → CRAG loop
                      → Graph enrich → Generate answer
                        └─ 🤖 Agentic mode: LLM picks tools autonomously

pip install agent-memory-mcp

{
  "mcpServers": {
    "agent-memory": {
      "command": "agent-memory-mcp",
      "env": {
        "AGENT_MEMORY_API_KEY": "amk_your_key_here",
        "AGENT_MEMORY_URL": "https://agent.ai-vfx.com"
      }
    }
  }
}

Endpoint: https://agent.ai-vfx.com/mcp
Auth: Bearer token (API key) or OAuth 2.0 with PKCE

# 🔍 Search memory
curl -X POST https://agent.ai-vfx.com/api/v1/memory/search \
  -H "Authorization: Bearer amk_your_key_here" \
  -H "Content-Type: application/json" \
  -d '{"query": "what decisions were made about wallet integration?"}'

# 📋 Get weekly digest
curl -X POST https://agent.ai-vfx.com/api/v1/digest \
  -H "Authorization: Bearer amk_your_key_here" \
  -H "Content-Type: application/json" \
  -d '{"scope": "@channel_name", "period": "7d"}'

# ✅ Get decisions
curl -X POST https://agent.ai-vfx.com/api/v1/decisions \
  -H "Authorization: Bearer amk_your_key_here" \
  -H "Content-Type: application/json" \
  -d '{"scope": "@team_chat", "topic": "release planning"}'

# 💰 Check balance
curl https://agent.ai-vfx.com/api/v1/account/balance \
  -H "Authorization: Bearer amk_your_key_here"

openclaw install may4vfx/telegram-agent-memory

📥 Collection       Telethon multi-user collector, encrypted sessions in DB
       ↓
🧹 Noise Filter     Remove joins/leaves, service messages, empty content
       ↓
📝 Metadata          Language detection, content type, timestamps
       ↓
🧵 Threading         Group replies into conversation threads, link forum topics
       ↓
🏷️ Entity Extract    LLM-based extraction of entities and relationships (parallel batches)
       ↓
🧬 Embedding         BGE-M3 dense vectors (1024-dim) via Text Embeddings Inference
       ↓
💾 Storage           Parallel write → PostgreSQL + Milvus + FalkorDB
       ↓
🕸️ Communities       Leiden algorithm clusters related entities in the graph
       ↓
🔍 Schema Discovery  Auto-detect domain type, entity types, relation types

Messages (up to 5000) → Engagement scoring (replies × 3 + content length)
    → Top-200 selection → BGE-M3 embedding → Cosine deduplication
    → Semantic clustering → Parallel LLM labeling (emoji + topic name)
    → MAP: summarize each cluster → REDUCE: synthesize final digest
    → Format with links to original Telegram messages