graph TB
    subgraph Clients["🖥️ Clients"]
        CLI["axel-chat / CLI"]
        WebUI["Open WebUI"]
        Discord["Discord Bot"]
        Telegram["Telegram Bot"]
    end

    subgraph Backend["⚙️ AXNMIHN Backend (FastAPI :8000)"]
        subgraph Ingress["Ingress Layer"]
            API["API Routers<br/>(REST/SSE/WebSocket)"]
            ChannelMgr["Channel Manager<br/>(Adapter Lifecycle)"]
        end

        subgraph Core["Core Services"]
            ChatHandler["ChatHandler<br/>(Orchestrator)"]
            Context["ContextService<br/>(Budget-aware)"]
            ReAct["ReActLoopService<br/>(≤15 iterations)"]
            Search["SearchService<br/>(Tavily)"]
            ToolExec["ToolExecutionService"]
            MemPersist["MemoryPersistenceService"]
            Emotion["EmotionService"]
        end

        subgraph LLM["LLM Router (Circuit Breaker)"]
            Claude["Anthropic<br/>Claude Sonnet 4.5<br/>(Chat)"]
            Gemini["Google<br/>Gemini 3 Flash<br/>(Utility)"]
        end

        subgraph Memory["6-Layer Memory System"]
            M0["M0: Event Buffer"]
            M1["M1: Working Memory<br/>(20 turns, JSON)"]
            M3["M3: Session Archive<br/>(PostgreSQL/SQLite)"]
            M4["M4: Long-Term<br/>(ChromaDB/pgvector)"]
            M51["M5.1: MemGPT<br/>(budget selection)"]
            M52["M5.2: GraphRAG<br/>(knowledge graph)"]
            M53["M5.3: MetaMemory<br/>(access patterns)"]
        end

        subgraph MCP["MCP Server (:8555)"]
            MCPTools["Tool Registry<br/>(32 tools)"]
        end

        subgraph Media["Media Pipeline"]
            TTS["TTS<br/>(Qwen3 / OpenAI)"]
            STT["STT<br/>(Deepgram Nova-3)"]
        end
    end

    subgraph External["🔌 External Services"]
        HASS["Home Assistant<br/>(WiZ/IoT)"]
        Playwright["Playwright<br/>(Browser)"]
        TavilyAPI["Tavily API"]
        DDGAPI["DuckDuckGo API"]
        PG["PostgreSQL"]
        Redis["Redis"]
    end

    CLI & WebUI -->|"OpenAI-compat API"| API
    Discord & Telegram -->|"Channel Adapter Protocol"| ChannelMgr
    ChannelMgr --> ChatHandler
    API --> ChatHandler

    ChatHandler --> Context
    ChatHandler --> ReAct
    ChatHandler --> Search
    ChatHandler --> ToolExec
    ChatHandler --> MemPersist
    ChatHandler --> Emotion

    Context --> Memory
    ReAct --> LLM
    ToolExec --> MCP
    MemPersist --> Memory
    ChatHandler --> Media

    MCPTools --> HASS
    MCPTools --> Playwright
    Search --> TavilyAPI
    Search --> DDGAPI

    M3 & M4 & M52 --> PG
    M0 --> M1 --> M3 -->|"consolidation (6h)"| M4
    M1 -->|"auto-promote"| M4
    M4 --> M51 & M52 & M53
    M52 -->|"feedback"| M4
    M53 -->|"hot boost"| M4

sequenceDiagram
    participant C as Client / Channel Bot
    participant A as API / ChannelManager
    participant CH as ChatHandler
    participant CS as ContextService
    participant MM as MemoryManager
    participant RL as ReActLoop
    participant LLM as LLM Router
    participant MCP as MCP Tools

    C->>A: Message (REST/WebSocket/Bot)
    A->>CH: ChatRequest
    CH->>CS: build_smart_context()
    CS->>MM: parallel fetch (M0+M1+M3+M4+M5)
    MM-->>CS: budgeted context
    CS-->>CH: assembled context

    loop ReAct Loop (≤15)
        CH->>RL: reason + act
        RL->>LLM: generate (Claude/Gemini)
        LLM-->>RL: response + tool_calls
        opt Tool Calls
            RL->>MCP: execute tools
            MCP-->>RL: tool results
        end
    end

    CH->>MM: persist (working + session + long-term)
    MM->>LLM: evaluate importance (facts/insights)
    LLM-->>MM: importance scores
    opt importance ≥ 0.6
        MM->>MM: auto-promote to M4
    end
    CH-->>A: streaming response (SSE)
    A-->>C: chunked response

graph TB
    Input["User Message"] --> M0

    subgraph Memory["6-Layer Memory System"]
        M0["M0: Event Buffer<br/><i>real-time event stream</i>"]
        M1["M1: Working Memory<br/><i>in-memory deque (20 turns)</i><br/><i>JSON persistence</i>"]
        M3["M3: Session Archive<br/><i>SQLite / PostgreSQL</i><br/><i>sessions, messages, logs</i>"]
        M4["M4: Long-Term Memory<br/><i>ChromaDB / pgvector</i><br/><i>3072-dim Gemini embeddings</i><br/><i>adaptive decay, dedup</i>"]

        M51["M5.1: MemGPT<br/><i>budget-aware selection</i><br/><i>token-budgeted assembly</i>"]
        M52["M5.2: GraphRAG<br/><i>entity-relation graph</i><br/><i>spaCy NER + LLM extraction</i><br/><i>BFS traversal (C++ for 100+)</i>"]
        M53["M5.3: MetaMemory<br/><i>access pattern tracking</i><br/><i>hot memory detection</i>"]
    end

    M0 --> M1
    M1 -->|"immediate persist"| M3
    M1 -->|"auto-promote<br/>(importance ≥ 0.6)"| M4
    M3 -->|"consolidation (6h)<br/>+ reassessment"| M4
    M4 --> M51
    M4 --> M52
    M4 --> M53
    M52 -->|"connection_count<br/>feedback"| M4
    M53 -->|"hot boost"| M4

    Context["ContextService<br/>build_smart_context()"] -.->|"parallel fetch"| M0 & M1 & M3 & M4 & M51 & M52 & M53

graph TB
    subgraph Clients["🖥️ 클라이언트"]
        CLI["axel-chat / CLI"]
        WebUI["Open WebUI"]
        Discord["Discord 봇"]
        Telegram["Telegram 봇"]
    end

    subgraph Backend["⚙️ AXNMIHN 백엔드 (FastAPI :8000)"]
        subgraph Ingress["인그레스 계층"]
            API["API 라우터<br/>(REST/SSE/WebSocket)"]
            ChannelMgr["채널 매니저<br/>(어댑터 라이프사이클)"]
        end

        subgraph Core["핵심 서비스"]
            ChatHandler["ChatHandler<br/>(오케스트레이터)"]
            Context["ContextService<br/>(예산 기반)"]
            ReAct["ReActLoopService<br/>(≤15 반복)"]
            Search["SearchService<br/>(Tavily)"]
            ToolExec["ToolExecutionService"]
            MemPersist["MemoryPersistenceService"]
            Emotion["EmotionService"]
        end

        subgraph LLM["LLM 라우터 (Circuit Breaker)"]
            Claude["Anthropic<br/>Claude Sonnet 4.5<br/>(채팅)"]
            Gemini["Google<br/>Gemini 3 Flash<br/>(유틸리티)"]
        end

        subgraph Memory["6계층 메모리 시스템"]
            M0["M0: 이벤트 버퍼"]
            M1["M1: 워킹 메모리<br/>(20턴, JSON)"]
            M3["M3: 세션 아카이브<br/>(PostgreSQL/SQLite)"]
            M4["M4: 장기 메모리<br/>(ChromaDB/pgvector)"]
            M51["M5.1: MemGPT<br/>(예산 기반 선택)"]
            M52["M5.2: GraphRAG<br/>(지식 그래프)"]
            M53["M5.3: MetaMemory<br/>(접근 패턴)"]
        end

        subgraph MCP["MCP 서버 (:8555)"]
            MCPTools["도구 레지스트리<br/>(32개 도구)"]
        end

        subgraph Media["미디어 파이프라인"]
            TTS["TTS<br/>(Qwen3 / OpenAI)"]
            STT["STT<br/>(Deepgram Nova-3)"]
        end
    end

    subgraph External["🔌 외부 서비스"]
        HASS["Home Assistant<br/>(WiZ/IoT)"]
        Playwright["Playwright<br/>(브라우저)"]
        TavilyAPI["Tavily API"]
        DDGAPI["DuckDuckGo API"]
        PG["PostgreSQL"]
        Redis["Redis"]
    end

    CLI & WebUI -->|"OpenAI 호환 API"| API
    Discord & Telegram -->|"채널 어댑터 프로토콜"| ChannelMgr
    ChannelMgr --> ChatHandler
    API --> ChatHandler

    ChatHandler --> Context
    ChatHandler --> ReAct
    ChatHandler --> Search
    ChatHandler --> ToolExec
    ChatHandler --> MemPersist
    ChatHandler --> Emotion

    Context --> Memory
    ReAct --> LLM
    ToolExec --> MCP
    MemPersist --> Memory
    ChatHandler --> Media

    MCPTools --> HASS
    MCPTools --> Playwright
    Search --> TavilyAPI
    Search --> DDGAPI

    M3 & M4 & M52 --> PG
    M0 --> M1 --> M3 -->|"통합 (6시간)"| M4
    M1 -->|"자동 프로모션"| M4
    M4 --> M51 & M52 & M53
    M52 -->|"피드백"| M4
    M53 -->|"핫 부스트"| M4

sequenceDiagram
    participant C as 클라이언트 / 채널 봇
    participant A as API / ChannelManager
    participant CH as ChatHandler
    participant CS as ContextService
    participant MM as MemoryManager
    participant RL as ReActLoop
    participant LLM as LLM 라우터
    participant MCP as MCP 도구

    C->>A: 메시지 (REST/WebSocket/봇)
    A->>CH: ChatRequest
    CH->>CS: build_smart_context()
    CS->>MM: 병렬 조회 (M0+M1+M3+M4+M5)
    MM-->>CS: 예산 기반 컨텍스트
    CS-->>CH: 조립된 컨텍스트

    loop ReAct 루프 (≤15회)
        CH->>RL: 추론 + 실행
        RL->>LLM: 생성 (Claude/Gemini)
        LLM-->>RL: 응답 + tool_calls
        opt 도구 호출
            RL->>MCP: 도구 실행
            MCP-->>RL: 도구 결과
        end
    end

    CH->>MM: 영속화 (working + session + long-term)
    MM->>LLM: 중요도 평가 (facts/insights)
    LLM-->>MM: importance 점수
    opt importance ≥ 0.6
        MM->>MM: M4로 자동 프로모션
    end
    CH-->>A: 스트리밍 응답 (SSE)
    A-->>C: 청크 응답

graph TB
    Input["사용자 메시지"] --> M0

    subgraph Memory["6계층 메모리 시스템"]
        M0["M0: 이벤트 버퍼<br/><i>실시간 이벤트 스트림</i>"]
        M1["M1: 워킹 메모리<br/><i>인메모리 deque (20턴)</i><br/><i>JSON 영속화</i>"]
        M3["M3: 세션 아카이브<br/><i>SQLite / PostgreSQL</i><br/><i>세션, 메시지, 로그</i>"]
        M4["M4: 장기 메모리<br/><i>ChromaDB / pgvector</i><br/><i>3072차원 Gemini 임베딩</i><br/><i>적응형 decay, 중복 제거</i>"]

        M51["M5.1: MemGPT<br/><i>예산 기반 선택</i><br/><i>토큰 예산 조립</i>"]
        M52["M5.2: GraphRAG<br/><i>엔티티-관계 그래프</i><br/><i>spaCy NER + LLM 추출</i><br/><i>BFS 탐색 (100+ 시 C++)</i>"]
        M53["M5.3: MetaMemory<br/><i>접근 패턴 추적</i><br/><i>핫 메모리 감지</i>"]
    end

    M0 --> M1
    M1 -->|"즉시 영속화"| M3
    M1 -->|"자동 프로모션<br/>(importance ≥ 0.6)"| M4
    M3 -->|"통합 (6시간)<br/>+ 재평가"| M4
    M4 --> M51
    M4 --> M52
    M4 --> M53
    M52 -->|"connection_count<br/>피드백"| M4
    M53 -->|"핫 부스트"| M4

    Context["ContextService<br/>build_smart_context()"] -.->|"병렬 조회"| M0 & M1 & M3 & M4 & M51 & M52 & M53