ai state machine init

.vscode/settings.json

@@ -0,0 +1,3 @@
+{
+    "java.configuration.updateBuildConfiguration": "automatic"
+}

app/note/design_doc

@@ -113,7 +113,85 @@ https://www.modelscope.cn/datasets/shaoxuan/WIDER_FACE/files
 8. 人脸识别模型是insightface的r18模型，转成了rknn格式，并且使用了 lfw 的数据集进行了校准，下载地址：
 https://tianchi.aliyun.com/dataset/93864
 
-9. 
+9. 本地LLM，用的是 rkllm 模型，由于内存限制，只能用较小的模型，如 Qwen3-0.6B-rk3588-w8a8.rkllm。
+
+10. 数字人交互设计说明书（优化版）
+核心原则
+* 尊重用户：给用户足够的反应时间，不抢话，不催促。
+* 主动但不打扰：用户沉默时适度引导，但不强行交互；数字人拥有独立的内心活动，即便无人交互也保持“生命感”。
+* 个性化：记住每个用户的偏好、名字、对话历史，并据此调整问候和回应。
+* 动作同步：Live2D人物的表情与动作与当前状态、情绪、内心活动相匹配。
+* 持续内心活动：数字人即使在没有用户时也会进行“回忆”或“思考”，并通过文字展现，用户可随时打断并询问想法。
+
+状态定义
+[空闲状态]：无人，数字人无内心活动（省电模式）。
+[回忆状态]：无人，数字人正在根据记忆产生内心想法，通过表情表现。
+[问候状态]：检测到用户，主动打招呼。
+[等待回复状态]：问候后或对话间隙，等待用户回复。
+[对话状态]：正在与用户交流，根据情绪调整动作。
+[主动引导状态]：用户沉默但仍在画面中，数字人主动开启新话题。
+[告别状态]：用户离开画面，数字人告别。
+[空闲状态]
+   动作 ： haru_g_idle （待机动作）
+    ↓
+检测到人脸 → 等待1秒（让人脸稳定）
+    ↓
+[问候状态]
+   动作 ：
+   - 认识用户： haru_g_m22  （高兴动作）
+   - 不认识： haru_g_m01（中性动作）
+   AI问候（根据是否认识个性化）
+   启动20秒计时器
+    ↓
+[等待回复状态]
+   动作 ： haru_g_m17 
+    ↓
+    ├─ 如果用户在20秒内回复 → 进入[对话状态]
+    │      动作 ：根据对话情绪动态调整
+    │      - 开心： haru_g_m22 / haru_g_m21 / haru_g_m18 / haru_g_m09 / haru_g_m08
+    │      - 悲伤： haru_g_m25 / haru_g_m24 / haru_g_m05 / haru_g_m16
+    │      - 惊讶： haru_g_m26 / haru_g_m12
+    │      - 愤怒： haru_g_m04 / haru_g_m11 / haru_g_m04
+    │      - 平静： haru_g_m15 / haru_g_m07 / haru_g_m06 / haru_g_m02 / haru_g_m01
+    │      AI根据内容回应，持续直到用户说“再见”
+    │      ↓
+    │      （用户离开）→ 进入[告别状态]
+    │
+    └─ 如果20秒内无回复
+            ↓
+        检查用户是否还在画面
+            ↓
+        ├─ 如果还在 → [主动引导状态]
+        │      动作 ： haru_g_m15 / haru_g_m07 （中性动作）
+        │      AI开启轻松话题（如“我出一道数学题考考你吧？1+6等于多少？”）
+        │      等待10秒
+        │          ↓
+        │      ├─ 回复 → 进入[对话状态]
+        │      └─ 没回复 → 换话题，如，我们上完厕所应该干什么呀？最多重复3次
+        │            动作 ： haru_g_m22 / haru_g_m18
+        │            （重复尝试）→ 再次等待10秒
+        │            （3次后仍无回复）→ 回到[等待回复状态]
+        │
+        └─ 如果已离开 → [告别状态]
+               动作 ： haru_g_idle （告别后待机）
+               ↓
+               3秒后 → 回到[空闲状态]
+
+[空闲状态] （长时间无人）
+    ↓
+   如果持续30秒无人 → 进入[回忆状态]
+       动作 ： haru_g_m15
+       回顾之前的聊天记录，生成想法，不发声，把想法显示在屏幕上
+       ↓
+       如果检测到人脸 → 立即中断回忆，进入[问候状态]
 
 
+[回忆状态] 被用户询问“你在想什么？”
+       ↓
+    AI说出最近的想法（如“我刚才在想，上次你说你喜欢蓝色...”）
+       ↓
+    进入[问候状态]或直接进入[对话状态]（根据用户后续反应）
+
+[告别状态] （用户离开后）
+   动作 ： 简短告别，用语音，→ 等待3秒 → 回到[空闲状态]
 

app/src/main/java/com/digitalperson/Live2DChatActivity.kt

@@ -27,10 +27,15 @@ import com.digitalperson.face.ImageProxyBitmapConverter
 import com.digitalperson.metrics.TraceManager
 import com.digitalperson.metrics.TraceSession
 import com.digitalperson.tts.TtsController
+import com.digitalperson.interaction.DigitalHumanInteractionController
+import com.digitalperson.interaction.InteractionActionHandler
+import com.digitalperson.interaction.InteractionState
+import com.digitalperson.interaction.UserMemoryStore
 import com.digitalperson.llm.LLMManager
 import com.digitalperson.llm.LLMManagerCallback
 import com.digitalperson.util.FileHelper
 import java.io.File
+import org.json.JSONObject
 import java.util.concurrent.ExecutorService
 import java.util.concurrent.Executors
 import kotlinx.coroutines.CoroutineScope
@@ -85,9 +90,18 @@ class Live2DChatActivity : AppCompatActivity() {
     private lateinit var cameraPreviewView: PreviewView
     private lateinit var faceOverlayView: FaceOverlayView
     private lateinit var faceDetectionPipeline: FaceDetectionPipeline
+    private lateinit var interactionController: DigitalHumanInteractionController
+    private lateinit var userMemoryStore: UserMemoryStore
     private var facePipelineReady: Boolean = false
     private var cameraProvider: ProcessCameraProvider? = null
     private lateinit var cameraAnalyzerExecutor: ExecutorService
+    private var activeUserId: String = "guest"
+    private var pendingLocalThoughtCallback: ((String) -> Unit)? = null
+    private var pendingLocalProfileCallback: ((String) -> Unit)? = null
+    private var localThoughtSilentMode: Boolean = false
+    private val recentConversationLines = ArrayList<String>()
+    private var recentConversationDirty: Boolean = false
+    private var lastFacePresent: Boolean = false
 
     override fun onRequestPermissionsResult(
         requestCode: Int,
@@ -143,15 +157,73 @@ class Live2DChatActivity : AppCompatActivity() {
         cameraPreviewView.implementationMode = PreviewView.ImplementationMode.COMPATIBLE
         faceOverlayView = findViewById(R.id.face_overlay)
         cameraAnalyzerExecutor = Executors.newSingleThreadExecutor()
+        userMemoryStore = UserMemoryStore(applicationContext)
+        interactionController = DigitalHumanInteractionController(
+            scope = ioScope,
+            handler = object : InteractionActionHandler {
+                override fun onStateChanged(state: InteractionState) {
+                    runOnUiThread {
+                        uiManager.appendToUi("\n[State] $state\n")
+                    }
+                    if (state == InteractionState.IDLE) {
+                        analyzeUserProfileInIdleIfNeeded()
+                    }
+                }
+
+                override fun playMotion(motionName: String) {
+                    playInteractionMotion(motionName)
+                }
+
+                override fun appendText(text: String) {
+                    uiManager.appendToUi(text)
+                }
+
+                override fun speak(text: String) {
+                    ttsController.enqueueSegment(text)
+                    ttsController.enqueueEnd()
+                }
+
+                override fun requestCloudReply(userText: String) {
+                    llmInFlight = true
+                    Log.i(TAG_LLM, "Routing dialogue to CLOUD")
+                    cloudApiManager.callLLM(buildCloudPromptWithUserProfile(userText))
+                }
+
+                override fun requestLocalThought(prompt: String, onResult: (String) -> Unit) {
+                    this@Live2DChatActivity.requestLocalThought(prompt, onResult)
+                }
+
+                override fun onRememberUser(faceIdentityId: String, name: String?) {
+                    activeUserId = faceIdentityId
+                    userMemoryStore.upsertUserSeen(activeUserId, name)
+                }
+
+                override fun saveThought(thought: String) {
+                    userMemoryStore.upsertUserSeen(activeUserId, null)
+                    userMemoryStore.updateThought(activeUserId, thought)
+                }
+
+                override fun loadLatestThought(): String? = userMemoryStore.getLatestThought()
+
+                override fun addToChatHistory(role: String, content: String) {
+                    appendConversationLine(role, content)
+                }
+
+                override fun addAssistantMessageToCloudHistory(content: String) {
+                    cloudApiManager.addAssistantMessage(content)
+                }
+            }
+        )
         faceDetectionPipeline = FaceDetectionPipeline(
             context = applicationContext,
             onResult = { result ->
                 faceOverlayView.updateResult(result)
             },
-            onGreeting = { greeting ->
+            onPresenceChanged = { present, faceIdentityId, recognizedName ->
-                uiManager.appendToUi("\n[Face] $greeting\n")
+                if (present == lastFacePresent) return@FaceDetectionPipeline
-                ttsController.enqueueSegment(greeting)
+                lastFacePresent = present
-                ttsController.enqueueEnd()
+                Log.d(TAG_ACTIVITY, "present=$present, faceIdentityId=$faceIdentityId, recognized=$recognizedName")
+                interactionController.onFacePresenceChanged(present, faceIdentityId, recognizedName)
             }
         )
         
@@ -200,13 +272,12 @@ class Live2DChatActivity : AppCompatActivity() {
 
         // 设置 LLM 模式开关
         uiManager.setLLMSwitchListener { isChecked ->
+            // 交互状态机固定路由：用户对话走云端，回忆走本地。此开关仅作为本地LLM可用性提示。
             useLocalLLM = isChecked
-            Log.i(TAG_LLM, "LLM mode switched: useLocalLLM=$useLocalLLM")
+            uiManager.showToast("状态机路由已固定：对话云端，回忆本地")
-            uiManager.showToast("LLM模式已切换到${if (isChecked) "本地" else "云端"}")
-            // 重新初始化 LLM
-            initLLM()
         }
         // 默认不显示 LLM 开关，等模型下载完成后再显示
+        uiManager.showLLMSwitch(false)
 
         if (AppConfig.USE_HOLD_TO_SPEAK) {
             uiManager.setButtonsEnabled(recordEnabled = false)
@@ -226,8 +297,9 @@ class Live2DChatActivity : AppCompatActivity() {
         vadManager = VadManager(this)
         vadManager.setCallback(createVadCallback())
         
-        // 初始化 LLM 管理器
+        // 初始化本地 LLM（用于 memory 状态）
         initLLM()
+        interactionController.start()
         
         // 检查是否需要下载模型
         if (!FileHelper.isLocalLLMAvailable(this)) {
@@ -259,8 +331,8 @@ class Live2DChatActivity : AppCompatActivity() {
                             if (FileHelper.isLocalLLMAvailable(this)) {
                                 Log.i(AppConfig.TAG, "Local LLM is available, enabling local LLM switch")
                                 // 显示本地 LLM 开关，并同步状态
-                                uiManager.showLLMSwitch(true)
+                                uiManager.showLLMSwitch(false)
-                                uiManager.setLLMSwitchChecked(useLocalLLM)
+                                initLLM()
                             }
                         } else {
                             Log.e(AppConfig.TAG, "Failed to download model files: $message")
@@ -275,8 +347,7 @@ class Live2DChatActivity : AppCompatActivity() {
             // 模型已存在，直接初始化其他组件
             initializeOtherComponents()
             // 显示本地 LLM 开关，并同步状态
-            uiManager.showLLMSwitch(true)
+            uiManager.showLLMSwitch(false)
-            uiManager.setLLMSwitchChecked(useLocalLLM)
         }
     }
     
@@ -348,6 +419,7 @@ class Live2DChatActivity : AppCompatActivity() {
             runOnUiThread {
                 uiManager.appendToUi("\n\n[ASR] ${text}\n")
             }
+            appendConversationLine("用户", text)
             currentTrace?.markRecordingDone()
             currentTrace?.markLlmResponseReceived()
         }
@@ -361,17 +433,8 @@ class Live2DChatActivity : AppCompatActivity() {
         override fun isLlmInFlight(): Boolean = llmInFlight
         
         override fun onLlmCalled(text: String) {
-            llmInFlight = true
+            Log.d(AppConfig.TAG, "Forward ASR text to interaction controller: $text")
-            Log.d(AppConfig.TAG, "Calling LLM with text: $text")
+            interactionController.onUserAsrText(text)
-            if (useLocalLLM) {
-                Log.i(TAG_LLM, "Routing to LOCAL LLM")
-                // 使用本地 LLM 生成回复
-                generateResponse(text)
-            } else {
-                Log.i(TAG_LLM, "Routing to CLOUD LLM")
-                // 使用云端 LLM 生成回复
-                cloudApiManager.callLLM(text)
-            }
         }
     }
     
@@ -390,6 +453,7 @@ class Live2DChatActivity : AppCompatActivity() {
         override fun onLLMResponseReceived(response: String) {
             currentTrace?.markLlmDone()
             llmInFlight = false
+            appendConversationLine("助手", response)
 
             if (enableStreaming) {
                 for (seg in segmenter.flush()) {
@@ -411,6 +475,7 @@ class Live2DChatActivity : AppCompatActivity() {
                 ttsController.enqueueSegment(filteredText)
                 ttsController.enqueueEnd()
             }
+            interactionController.onDialogueResponseFinished()
         }
 
         override fun onLLMStreamingChunkReceived(chunk: String) {
@@ -442,6 +507,7 @@ class Live2DChatActivity : AppCompatActivity() {
         override fun onError(errorMessage: String) {
             llmInFlight = false
             uiManager.showToast(errorMessage, Toast.LENGTH_LONG)
+            interactionController.onDialogueResponseFinished()
             onStopClicked(userInitiated = false)
         }
     }
@@ -479,6 +545,7 @@ class Live2DChatActivity : AppCompatActivity() {
 
     override fun onDestroy() {
         super.onDestroy()
+        try { interactionController.stop() } catch (_: Throwable) {}
         stopCameraPreviewAndDetection()
         onStopClicked(userInitiated = false)
         ioScope.cancel()
@@ -490,6 +557,7 @@ class Live2DChatActivity : AppCompatActivity() {
         try { cameraAnalyzerExecutor.shutdown() } catch (_: Throwable) {}
         try { ttsController.release() } catch (_: Throwable) {}
         try { llmManager?.destroy() } catch (_: Throwable) {}
+        try { userMemoryStore.close() } catch (_: Throwable) {}
         try { uiManager.release() } catch (_: Throwable) {}
         try { audioProcessor.release() } catch (_: Throwable) {}
     }
@@ -757,85 +825,158 @@ class Live2DChatActivity : AppCompatActivity() {
         }
         Log.d(AppConfig.TAG, "processSamplesLoop stopped")
     }
+
+    private fun playInteractionMotion(motionName: String) {
+        when (motionName) {
+            "haru_g_m22.motion3.json" -> uiManager.setMood("高兴")
+            "haru_g_m01.motion3.json", "haru_g_m17.motion3.json" -> uiManager.setMood("中性")
+            "haru_g_m15.motion3.json" -> uiManager.setMood("关心")
+            "haru_g_idle.motion3.json" -> uiManager.setMood("平和")
+            else -> uiManager.setMood("中性")
+        }
+    }
+
+    private fun appendConversationLine(role: String, text: String) {
+        val line = "$role: ${text.trim()}"
+        if (line.length <= 4) return
+        recentConversationLines.add(line)
+        if (recentConversationLines.size > 12) {
+            recentConversationLines.removeAt(0)
+        }
+        recentConversationDirty = true
+    }
+
+    private fun buildCloudPromptWithUserProfile(userText: String): String {
+        val profile = userMemoryStore.getMemory(activeUserId) ?: return userText
+        val profileParts = ArrayList<String>()
+        profile.displayName?.takeIf { it.isNotBlank() }?.let { profileParts.add("姓名:$it") }
+        profile.age?.takeIf { it.isNotBlank() }?.let { profileParts.add("年龄:$it") }
+        profile.gender?.takeIf { it.isNotBlank() }?.let { profileParts.add("性别:$it") }
+        profile.hobbies?.takeIf { it.isNotBlank() }?.let { profileParts.add("爱好:$it") }
+        profile.profileSummary?.takeIf { it.isNotBlank() }?.let { profileParts.add("画像:$it") }
+        if (profileParts.isEmpty()) return userText
+        return buildString {
+            append("[用户画像]\n")
+            append(profileParts.joinToString("；"))
+            append("\n[/用户画像]\n")
+            append(userText)
+        }
+    }
+
+    private fun analyzeUserProfileInIdleIfNeeded() {
+        if (!recentConversationDirty || !activeUserId.startsWith("face_")) return
+        if (recentConversationLines.isEmpty()) return
+        val dialogue = recentConversationLines.joinToString("\n")
+        requestLocalProfileExtraction(dialogue) { raw ->
+            try {
+                val json = parseFirstJsonObject(raw)
+                val name = json.optString("name", "").trim().ifBlank { null }
+                val age = json.optString("age", "").trim().ifBlank { null }
+                val gender = json.optString("gender", "").trim().ifBlank { null }
+                val hobbies = json.optString("hobbies", "").trim().ifBlank { null }
+                val summary = json.optString("summary", "").trim().ifBlank { null }
+                if (name != null) {
+                    userMemoryStore.updateDisplayName(activeUserId, name)
+                }
+                userMemoryStore.updateProfile(activeUserId, age, gender, hobbies, summary)
+                recentConversationDirty = false
+                runOnUiThread {
+                    uiManager.appendToUi("\n[Memory] 已更新用户画像: $activeUserId\n")
+                }
+            } catch (e: Exception) {
+                Log.w(TAG_LLM, "Profile parse failed: ${e.message}")
+            }
+        }
+    }
+
+    private fun requestLocalProfileExtraction(dialogue: String, onResult: (String) -> Unit) {
+        try {
+            val local = llmManager
+            if (local == null) {
+                onResult("{}")
+                return
+            }
+            localThoughtSilentMode = true
+            pendingLocalProfileCallback = onResult
+            Log.i(TAG_LLM, "Routing profile extraction to LOCAL")
+            local.generateResponseWithSystem(
+                "你是信息抽取器。仅输出JSON对象，不要其他文字。字段为name,age,gender,hobbies,summary。",
+                "请从以下对话提取用户信息，未知填空字符串：\n$dialogue"
+            )
+        } catch (e: Exception) {
+            pendingLocalProfileCallback = null
+            localThoughtSilentMode = false
+            Log.e(TAG_LLM, "requestLocalProfileExtraction failed: ${e.message}", e)
+            onResult("{}")
+        }
+    }
+
+    private fun parseFirstJsonObject(text: String): JSONObject {
+        val raw = text.trim()
+        val start = raw.indexOf('{')
+        val end = raw.lastIndexOf('}')
+        if (start >= 0 && end > start) {
+            return JSONObject(raw.substring(start, end + 1))
+        }
+        return JSONObject(raw)
+    }
     
     /**
-     * 初始化 LLM 管理器
+     * 初始化本地 LLM（仅用于回忆状态）
      */
     private fun initLLM() {
         try {
-            Log.i(TAG_LLM, "initLLM called, useLocalLLM=$useLocalLLM")
+            Log.i(TAG_LLM, "initLLM called for memory-local model")
             llmManager?.destroy()
             llmManager = null
-            if (useLocalLLM) {
+            val modelPath = FileHelper.getLLMModelPath(applicationContext)
-                // // 本地 LLM 初始化前，先暂停/释放重模块
+            if (!File(modelPath).exists()) {
-                // Log.i(AppConfig.TAG, "Pausing camera and releasing face detection before LLM initialization")
+                throw IllegalStateException("RKLLM model file missing: $modelPath")
-                // stopCameraPreviewAndDetection()
-                // try {
-                //     faceDetectionPipeline.release()
-                //     Log.i(AppConfig.TAG, "Face detection pipeline released")
-                // } catch (e: Exception) {
-                //     Log.w(AppConfig.TAG, "Failed to release face detection pipeline: ${e.message}")
-                // }
-                
-                // // 释放 VAD 管理器
-                // try {
-                //     vadManager.release()
-                //     Log.i(AppConfig.TAG, "VAD manager released")
-                // } catch (e: Exception) {
-                //     Log.w(AppConfig.TAG, "Failed to release VAD manager: ${e.message}")
-                // }
-                
-                val modelPath = FileHelper.getLLMModelPath(applicationContext)
-                if (!File(modelPath).exists()) {
-                    throw IllegalStateException("RKLLM model file missing: $modelPath")
-                }
-                Log.i(AppConfig.TAG, "Initializing LLM with model path: $modelPath")
-                val localLlmResponseBuffer = StringBuilder()
-                llmManager = LLMManager(modelPath, object : LLMManagerCallback {
-                    override fun onThinking(msg: String, finished: Boolean) {
-                        // 处理思考过程
-                        Log.d(TAG_LLM, "LOCAL onThinking finished=$finished msg=${msg.take(60)}")
-                        runOnUiThread {
-                            if (!finished && enableStreaming) {
-                                uiManager.appendToUi("\n[LLM] 思考中: $msg\n")
-                            }
-                        }
-                    }
-
-                    override fun onResult(msg: String, finished: Boolean) {
-                        // 处理生成结果
-                        Log.d(TAG_LLM, "LOCAL onResult finished=$finished len=${msg.length}")
-                        runOnUiThread {
-                            if (!finished) {
-                                localLlmResponseBuffer.append(msg)
-                                if (enableStreaming) {
-                                    uiManager.appendToUi(msg)
-                                }
-                            } else {
-                                val finalText = localLlmResponseBuffer.toString().trim()
-                                localLlmResponseBuffer.setLength(0)
-                                if (!enableStreaming && finalText.isNotEmpty()) {
-                                    uiManager.appendToUi("$finalText\n")
-                                }
-                                uiManager.appendToUi("\n\n[LLM] 生成完成\n")
-                                llmInFlight = false
-                                if (finalText.isNotEmpty()) {
-                                    ttsController.enqueueSegment(finalText)
-                                    ttsController.enqueueEnd()
-                                } else {
-                                    Log.w(TAG_LLM, "LOCAL final text is empty, skip TTS enqueue")
-                                }
-                            }
-                        }
-                    }
-                })
-                Log.i(AppConfig.TAG, "LLM initialized successfully")
-                Log.i(TAG_LLM, "LOCAL LLM initialized")
-            } else {
-                // 使用云端 LLM，不需要初始化本地 LLM
-                Log.i(AppConfig.TAG, "Using cloud LLM, skipping local LLM initialization")
-                Log.i(TAG_LLM, "CLOUD mode active")
             }
+            Log.i(AppConfig.TAG, "Initializing local memory LLM with model path: $modelPath")
+            val localLlmResponseBuffer = StringBuilder()
+            llmManager = LLMManager(modelPath, object : LLMManagerCallback {
+                override fun onThinking(msg: String, finished: Boolean) {
+                    Log.d(TAG_LLM, "LOCAL onThinking finished=$finished msg=${msg.take(60)}")
+                }
+
+                override fun onResult(msg: String, finished: Boolean) {
+                    Log.d(TAG_LLM, "LOCAL onResult finished=$finished len=${msg.length}")
+                    runOnUiThread {
+                        if (!finished) {
+                            localLlmResponseBuffer.append(msg)
+                            if (enableStreaming && !localThoughtSilentMode) {
+                                uiManager.appendToUi(msg)
+                            }
+                            return@runOnUiThread
+                        }
+                        val finalText = localLlmResponseBuffer.toString().trim()
+                        localLlmResponseBuffer.setLength(0)
+                        val profileCallback = pendingLocalProfileCallback
+                        pendingLocalProfileCallback = null
+                        if (profileCallback != null) {
+                            profileCallback(finalText)
+                            localThoughtSilentMode = false
+                            return@runOnUiThread
+                        }
+                        val callback = pendingLocalThoughtCallback
+                        pendingLocalThoughtCallback = null
+                        if (callback != null) {
+                            callback(finalText)
+                            localThoughtSilentMode = false
+                            return@runOnUiThread
+                        }
+                        if (!localThoughtSilentMode && finalText.isNotEmpty()) {
+                            uiManager.appendToUi("$finalText\n")
+                            ttsController.enqueueSegment(finalText)
+                            ttsController.enqueueEnd()
+                        }
+                        localThoughtSilentMode = false
+                    }
+                }
+            })
+            Log.i(TAG_LLM, "LOCAL memory LLM initialized")
+            useLocalLLM = true
         } catch (e: Exception) {
             Log.e(AppConfig.TAG, "Failed to initialize LLM: ${e.message}", e)
             Log.e(TAG_LLM, "LOCAL init failed: ${e.message}", e)
@@ -849,35 +990,27 @@ class Live2DChatActivity : AppCompatActivity() {
     }
     
     /**
-     * 使用 LLM 生成回复
+     * 回忆状态调用本地 LLM，仅用于 memory/what-are-you-thinking
      */
-    private fun generateResponse(userInput: String) {
+    private fun requestLocalThought(prompt: String, onResult: (String) -> Unit) {
         try {
-            if (useLocalLLM) {
+            val local = llmManager
-                val systemPrompt = "你是一个友好的数字人助手，回答要简洁明了。"
+            if (local == null) {
-                Log.d(AppConfig.TAG, "Generating response for: $userInput")
+                onResult("我在想，下次见面可以聊聊今天的新鲜事。")
-                val local = llmManager
+                return
-                if (local == null) {
-                    Log.e(TAG_LLM, "LOCAL LLM manager is null, fallback to CLOUD")
-                    cloudApiManager.callLLM(userInput)
-                    return
-                }
-                Log.i(TAG_LLM, "LOCAL generateResponseWithSystem")
-                local.generateResponseWithSystem(systemPrompt, userInput)
-            } else {
-                // 使用云端 LLM
-                Log.d(AppConfig.TAG, "Using cloud LLM for response: $userInput")
-                Log.i(TAG_LLM, "CLOUD callLLM")
-                // 调用云端 LLM
-                cloudApiManager.callLLM(userInput)
             }
+            localThoughtSilentMode = true
+            pendingLocalThoughtCallback = onResult
+            Log.i(TAG_LLM, "Routing memory thought to LOCAL")
+            local.generateResponseWithSystem(
+                "你是数字人内心独白模块，输出一句简短温和的想法。",
+                prompt
+            )
         } catch (e: Exception) {
-            Log.e(AppConfig.TAG, "Failed to generate response: ${e.message}", e)
+            Log.e(TAG_LLM, "requestLocalThought failed: ${e.message}", e)
-            Log.e(TAG_LLM, "generateResponse failed: ${e.message}", e)
+            pendingLocalThoughtCallback = null
-            runOnUiThread {
+            localThoughtSilentMode = false
-                uiManager.appendToUi("\n\n[Error] LLM 生成失败: ${e.message}\n")
+            onResult("我在想，下次见面可以聊聊今天的新鲜事。")
-                llmInFlight = false
-            }
         }
     }
 }

app/src/main/java/com/digitalperson/cloud/CloudApiManager.java

@@ -241,6 +241,13 @@ public class CloudApiManager {
         mConversationHistory = new JSONArray();
     }
     
+    /**
+     * 添加助手消息到对话历史
+     */
+    public void addAssistantMessage(String content) {
+        addMessageToHistory("assistant", content);
+    }
+    
     public void callTTS(String text, File outputFile) {
         if (mListener != null) {
             mMainHandler.post(() -> {

app/src/main/java/com/digitalperson/config/AppConfig.kt

@@ -73,4 +73,21 @@ object AppConfig {
         const val SECRET_ID = "AKIDbBdyBGE5oPuIGA1iDlDYlFallaJ0YODB" // 替换为你的腾讯云SECRET_ID
         const val SECRET_KEY = "32vhIl9OQIRclmLjvuleLp9LLAnFVYEp" // 替换为你的腾讯云SECRET_KEY
     }
+    
+    object LLM {
+        // 模型下载服务器地址
+        const val DOWNLOAD_SERVER = "http://192.168.1.19:5000"
+        // 下载路径
+        const val DOWNLOAD_PATH = "/download"
+        // 模型文件名
+        const val MODEL_FILE_NAME = "Qwen3-0.6B-rk3588-w8a8.rkllm"
+        // 模型存储目录
+        const val MODEL_DIR = "llm"
+        // 下载连接超时（毫秒）
+        const val DOWNLOAD_CONNECT_TIMEOUT = 600000
+        // 下载读取超时（毫秒）
+        const val DOWNLOAD_READ_TIMEOUT = 1200000
+        // 模型文件大小估计（字节）
+        const val MODEL_SIZE_ESTIMATE = 500L * 1024 * 1024 // 500MB
+    }
 }

app/src/main/java/com/digitalperson/face/FaceDetectionPipeline.kt

@@ -31,7 +31,7 @@ data class FaceDetectionResult(
 class FaceDetectionPipeline(
     private val context: Context,
     private val onResult: (FaceDetectionResult) -> Unit,
-    private val onGreeting: (String) -> Unit,
+    private val onPresenceChanged: (present: Boolean, faceIdentityId: String?, recognizedName: String?) -> Unit,
 ) {
     private val engine = RetinaFaceEngineRKNN()
     private val recognizer = FaceRecognizer(context)
@@ -41,8 +41,9 @@ class FaceDetectionPipeline(
     private var trackFace: FaceBox? = null
     private var trackId: Long = 0
     private var trackStableSinceMs: Long = 0
-    private var greetedTrackId: Long = -1
+    private var analyzedTrackId: Long = -1
-    private var lastGreetMs: Long = 0
+    private var lastFaceIdentityId: String? = null
+    private var lastRecognizedName: String? = null
 
     fun initialize(): Boolean {
         val detectorOk = engine.initialize(context)
@@ -98,8 +99,9 @@ class FaceDetectionPipeline(
 //                    )
 //                }
 
-                maybeRecognizeAndGreet(bitmap, filteredFaces)
+                maybeRecognize(bitmap, filteredFaces)
                 withContext(Dispatchers.Main) {
+                    onPresenceChanged(filteredFaces.isNotEmpty(), lastFaceIdentityId, lastRecognizedName)
                     onResult(FaceDetectionResult(width, height, filteredFaces))
                 }
             } catch (t: Throwable) {
@@ -111,11 +113,14 @@ class FaceDetectionPipeline(
         }
     }
 
-    private suspend fun maybeRecognizeAndGreet(bitmap: Bitmap, faces: List<FaceBox>) {
+    private fun maybeRecognize(bitmap: Bitmap, faces: List<FaceBox>) {
         val now = System.currentTimeMillis()
         if (faces.isEmpty()) {
             trackFace = null
             trackStableSinceMs = 0
+            analyzedTrackId = -1
+            lastFaceIdentityId = null
+            lastRecognizedName = null
             return
         }
 
@@ -123,62 +128,30 @@ class FaceDetectionPipeline(
         val prev = trackFace
         if (prev == null || iou(prev, primary) < AppConfig.Face.TRACK_IOU_THRESHOLD) {
             trackId += 1
-            greetedTrackId = -1
             trackStableSinceMs = now
+            analyzedTrackId = -1
+            lastFaceIdentityId = null
+            lastRecognizedName = null
         }
         trackFace = primary
 
         val stableMs = now - trackStableSinceMs
         val frontal = isFrontal(primary, bitmap.width, bitmap.height)
-        val coolingDown = (now - lastGreetMs) < AppConfig.FaceRecognition.GREETING_COOLDOWN_MS
+        if (stableMs < AppConfig.Face.STABLE_MS || !frontal) {
-        if (stableMs < AppConfig.Face.STABLE_MS || !frontal || greetedTrackId == trackId || coolingDown) {
+            return
+        }
+        if (analyzedTrackId == trackId) {
             return
         }
 
-        val match = recognizer.identify(bitmap, primary)
+        val match = recognizer.resolveIdentity(bitmap, primary)
-        
+        analyzedTrackId = trackId
-        Log.d(AppConfig.TAG, "[Face] Recognition result: matchedName=${match.matchedName}, similarity=${match.similarity}")
+        lastFaceIdentityId = match.matchedId?.let { "face_$it" }
-        
+        lastRecognizedName = match.matchedName
-        // 检查是否需要保存新人脸
-        if (match.matchedName.isNullOrBlank()) {
-            Log.d(AppConfig.TAG, "[Face] No match found, attempting to add new face")
-            // 提取人脸特征
-            val embedding = extractEmbedding(bitmap, primary)
-            Log.d(AppConfig.TAG, "[Face] Extracted embedding size: ${embedding.size}")
-            if (embedding.isNotEmpty()) {
-                // 尝试添加新人脸
-                val added = recognizer.addNewFace(embedding)
-                Log.d(AppConfig.TAG, "[Face] Add new face result: $added")
-                if (added) {
-                    Log.i(AppConfig.TAG, "[Face] New face added to database")
-                } else {
-                    Log.i(AppConfig.TAG, "[Face] Face already exists in database (similar face found)")
-                }
-            } else {
-                Log.w(AppConfig.TAG, "[Face] Failed to extract embedding")
-            }
-        } else {
-            Log.d(AppConfig.TAG, "[Face] Matched existing face: ${match.matchedName}")
-        }
-        
-        val greeting = if (!match.matchedName.isNullOrBlank()) {
-            "你好，${match.matchedName}！"
-        } else {
-            "你好，很高兴见到你。"
-        }
-        greetedTrackId = trackId
-        lastGreetMs = now
         Log.i(
             AppConfig.TAG,
-            "[Face] greeting track=$trackId stable=${stableMs}ms frontal=$frontal matched=${match.matchedName} score=${match.similarity}"
+            "[Face] stable track=$trackId faceId=${lastFaceIdentityId} matched=${match.matchedName} score=${match.similarity}"
         )
-        withContext(Dispatchers.Main) {
-            onGreeting(greeting)
-        }
-    }
-    
-    private fun extractEmbedding(bitmap: Bitmap, face: FaceBox): FloatArray {
-        return recognizer.extractEmbedding(bitmap, face)
     }
 
     private fun isFrontal(face: FaceBox, frameW: Int, frameH: Int): Boolean {

app/src/main/java/com/digitalperson/face/FaceFeatureStore.kt

@@ -11,7 +11,7 @@ import java.nio.ByteOrder
 
 data class FaceProfile(
     val id: Long,
-    val name: String,
+    val name: String?,
     val embedding: FloatArray,
 )
 
@@ -21,7 +21,7 @@ class FaceFeatureStore(context: Context) : SQLiteOpenHelper(context, DB_NAME, nu
             """
             CREATE TABLE IF NOT EXISTS face_profiles (
                 id INTEGER PRIMARY KEY AUTOINCREMENT,
-                name TEXT NOT NULL UNIQUE,
+                name TEXT,
                 embedding BLOB NOT NULL,
                 updated_at INTEGER NOT NULL
             )
@@ -55,9 +55,8 @@ class FaceFeatureStore(context: Context) : SQLiteOpenHelper(context, DB_NAME, nu
         return list
     }
 
-    fun upsertProfile(name: String, embedding: FloatArray) {
+    fun insertProfile(name: String?, embedding: FloatArray): Long {
-        // 确保名字不为null，使用空字符串作为默认值
+        val safeName = name?.takeIf { it.isNotBlank() }
-        val safeName = name.takeIf { it.isNotBlank() } ?: ""
         val values = ContentValues().apply {
             put("name", safeName)
             put("embedding", floatArrayToBlob(embedding))
@@ -67,9 +66,10 @@ class FaceFeatureStore(context: Context) : SQLiteOpenHelper(context, DB_NAME, nu
             "face_profiles",
             null,
             values,
-            SQLiteDatabase.CONFLICT_REPLACE
+            SQLiteDatabase.CONFLICT_NONE
         )
-        Log.i(AppConfig.TAG, "[FaceFeatureStore] upsertProfile name='$safeName' rowId=$rowId dim=${embedding.size}")
+        Log.i(AppConfig.TAG, "[FaceFeatureStore] insertProfile name='$safeName' rowId=$rowId dim=${embedding.size}")
+        return rowId
     }
 
     private fun floatArrayToBlob(values: FloatArray): ByteArray {
@@ -88,6 +88,6 @@ class FaceFeatureStore(context: Context) : SQLiteOpenHelper(context, DB_NAME, nu
 
     companion object {
         private const val DB_NAME = "face_feature.db"
-        private const val DB_VERSION = 1
+        private const val DB_VERSION = 2
     }
 }

app/src/main/java/com/digitalperson/face/FaceRecognizer.kt

@@ -8,6 +8,7 @@ import com.digitalperson.engine.ArcFaceEngineRKNN
 import kotlin.math.sqrt
 
 data class FaceRecognitionResult(
+    val matchedId: Long?,
     val matchedName: String?,
     val similarity: Float,
     val embeddingDim: Int,
@@ -40,10 +41,11 @@ class FaceRecognizer(context: Context) {
     }
 
     fun identify(bitmap: Bitmap, face: FaceBox): FaceRecognitionResult {
-        if (!initialized) return FaceRecognitionResult(null, 0f, 0)
+        if (!initialized) return FaceRecognitionResult(null, null, 0f, 0)
         val embedding = extractEmbedding(bitmap, face)
-        if (embedding.isEmpty()) return FaceRecognitionResult(null, 0f, 0)
+        if (embedding.isEmpty()) return FaceRecognitionResult(null, null, 0f, 0)
 
+        var bestId: Long? = null
         var bestName: String? = null
         var bestScore = -1f
         for (p in cache) {
@@ -51,13 +53,14 @@ class FaceRecognizer(context: Context) {
             val score = cosineSimilarity(embedding, p.embedding)
             if (score > bestScore) {
                 bestScore = score
+                bestId = p.id
                 bestName = p.name
             }
         }
         if (bestScore >= AppConfig.FaceRecognition.SIMILARITY_THRESHOLD) {
-            return FaceRecognitionResult(bestName, bestScore, embedding.size)
+            return FaceRecognitionResult(bestId, bestName, bestScore, embedding.size)
         }
-        return FaceRecognitionResult(null, bestScore, embedding.size)
+        return FaceRecognitionResult(null, null, bestScore, embedding.size)
     }
     
     fun extractEmbedding(bitmap: Bitmap, face: FaceBox): FloatArray {
@@ -65,38 +68,28 @@ class FaceRecognizer(context: Context) {
         return engine.extractEmbedding(bitmap, face.left, face.top, face.right, face.bottom)
     }
 
-    fun addOrUpdateProfile(name: String?, embedding: FloatArray) {
+    private fun addProfile(name: String?, embedding: FloatArray): Long {
         val normalized = normalize(embedding)
-        store.upsertProfile(name ?: "", normalized)
+        val rowId = store.insertProfile(name, normalized)
-        // 移除旧的记录（如果存在）
+        if (rowId > 0) {
-        if (name != null) {
+            cache.add(FaceProfile(id = rowId, name = name, embedding = normalized))
-            cache.removeAll { it.name == name }
         }
-        cache.add(FaceProfile(id = -1L, name = name ?: "", embedding = normalized))
+        return rowId
     }
 
-    fun addNewFace(embedding: FloatArray): Boolean {
+    fun resolveIdentity(bitmap: Bitmap, face: FaceBox): FaceRecognitionResult {
-        Log.d(AppConfig.TAG, "[FaceRecognizer] addNewFace: embedding size=${embedding.size}, cache size=${cache.size}")
+        val match = identify(bitmap, face)
-        
+        if (match.matchedId != null) return match
-        // 检查是否已经存在相似的人脸
+        val embedding = extractEmbedding(bitmap, face)
-        for (p in cache) {
+        if (embedding.isEmpty()) return match
-            if (p.embedding.size != embedding.size) {
+        val newId = addProfile(name = null, embedding = embedding)
-                Log.d(AppConfig.TAG, "[FaceRecognizer] Skipping profile with different embedding size: ${p.embedding.size}")
+        if (newId <= 0L) return match
-                continue
+        return FaceRecognitionResult(
-            }
+            matchedId = newId,
-            val score = cosineSimilarity(embedding, p.embedding)
+            matchedName = null,
-            Log.d(AppConfig.TAG, "[FaceRecognizer] Comparing with profile '${p.name}': similarity=$score, threshold=${AppConfig.FaceRecognition.SIMILARITY_THRESHOLD}")
+            similarity = match.similarity,
-            if (score >= AppConfig.FaceRecognition.SIMILARITY_THRESHOLD) {
+            embeddingDim = embedding.size
-                // 已经存在相似的人脸，不需要添加
+        )
-                Log.i(AppConfig.TAG, "[FaceRecognizer] Similar face found: ${p.name} with similarity=$score, not adding new face")
-                return false
-            }
-        }
-        
-        // 添加新人脸，名字为null
-        Log.i(AppConfig.TAG, "[FaceRecognizer] No similar face found, adding new face")
-        addOrUpdateProfile(null, embedding)
-        return true
     }
 
     fun release() {

app/src/main/java/com/digitalperson/interaction/DigitalHumanInteractionController.kt

@@ -0,0 +1,268 @@
+package com.digitalperson.interaction
+
+import android.util.Log
+import com.digitalperson.config.AppConfig
+import kotlinx.coroutines.CoroutineScope
+import kotlinx.coroutines.Job
+import kotlinx.coroutines.delay
+import kotlinx.coroutines.launch
+
+enum class InteractionState {
+    IDLE,
+    MEMORY,
+    GREETING,
+    WAITING_REPLY,
+    DIALOGUE,
+    PROACTIVE,
+    FAREWELL,
+}
+
+enum class LlmRoute {
+    CLOUD,
+    LOCAL,
+}
+
+interface InteractionActionHandler {
+    fun onStateChanged(state: InteractionState)
+    fun playMotion(motionName: String)
+    fun appendText(text: String)
+    fun speak(text: String)
+    fun requestCloudReply(userText: String)
+    fun requestLocalThought(prompt: String, onResult: (String) -> Unit)
+    fun onRememberUser(faceIdentityId: String, name: String?)
+    fun saveThought(thought: String)
+    fun loadLatestThought(): String?
+    fun addToChatHistory(role: String, content: String)
+    fun addAssistantMessageToCloudHistory(content: String)
+}
+
+class DigitalHumanInteractionController(
+    private val scope: CoroutineScope,
+    private val handler: InteractionActionHandler,
+) {
+    private val TAG: String = "DigitalHumanInteraction"
+    private var state: InteractionState = InteractionState.IDLE
+    private var facePresent: Boolean = false
+    private var recognizedName: String? = null
+    private var proactiveRound = 0
+    private var hasPendingUserReply = false
+
+    private var faceStableJob: Job? = null
+    private var waitReplyJob: Job? = null
+    private var proactiveJob: Job? = null
+    private var memoryJob: Job? = null
+    private var farewellJob: Job? = null
+
+    fun start() {
+        transitionTo(InteractionState.IDLE)
+        scheduleMemoryMode()
+    }
+
+fun onFacePresenceChanged(present: Boolean, faceIdentityId: String?, recognized: String?) {
+    Log.d(TAG, "onFacePresenceChanged: present=$present, faceIdentityId=$faceIdentityId, recognized=$recognized, state=$state")
+    
+    facePresent = present
+    if (!faceIdentityId.isNullOrBlank()) {
+        handler.onRememberUser(faceIdentityId, recognized)
+    }
+    if (!recognized.isNullOrBlank()) {
+        recognizedName = recognized
+    }
+
+    // 统一延迟处理
+    faceStableJob?.cancel()
+    faceStableJob = scope.launch {
+        delay(AppConfig.Face.STABLE_MS)
+        
+        if (present) {
+            // 人脸出现后的处理
+            if (facePresent && (state == InteractionState.IDLE || state == InteractionState.MEMORY)) {
+                enterGreeting()
+            } else if (state == InteractionState.FAREWELL) {
+                enterGreeting()
+            }
+        } else {
+            // 人脸消失后的处理
+            if (state != InteractionState.IDLE && state != InteractionState.MEMORY && state != InteractionState.FAREWELL) {
+                enterFarewell()
+            } else {
+                scheduleMemoryMode()
+            }
+        }
+    }
+}
+
+    fun onUserAsrText(text: String) {
+        val userText = text.trim()
+        if (userText.isBlank()) return
+
+        if (userText.contains("你在想什么")) {
+            val thought = handler.loadLatestThought()
+            if (!thought.isNullOrBlank()) {
+                handler.speak("我刚才在想：$thought")
+                handler.appendText("\n[回忆] $thought\n")
+                transitionTo(InteractionState.DIALOGUE)
+                handler.playMotion("haru_g_m15.motion3.json")
+                return
+            }
+        }
+
+        if (userText.contains("再见")) {
+            enterFarewell()
+            return
+        }
+
+        hasPendingUserReply = true
+        when (state) {
+            InteractionState.WAITING_REPLY, InteractionState.PROACTIVE, InteractionState.GREETING, InteractionState.DIALOGUE -> {
+                transitionTo(InteractionState.DIALOGUE)
+                requestDialogueReply(userText)
+            }
+            InteractionState.MEMORY, InteractionState.IDLE -> {
+                transitionTo(InteractionState.DIALOGUE)
+                requestDialogueReply(userText)
+            }
+            InteractionState.FAREWELL -> {
+                enterGreeting()
+            }
+        }
+    }
+
+    fun onDialogueResponseFinished() {
+        if (!facePresent) {
+            enterFarewell()
+            return
+        }
+        transitionTo(InteractionState.WAITING_REPLY)
+        handler.playMotion("haru_g_m17.motion3.json")
+        scheduleWaitingReplyTimeout()
+    }
+
+    private fun enterGreeting() {
+        transitionTo(InteractionState.GREETING)
+        val greet = if (!recognizedName.isNullOrBlank()) {
+            handler.playMotion("haru_g_m22.motion3.json")
+            "你好，$recognizedName，很高兴再次见到你。"
+        } else {
+            handler.playMotion("haru_g_m01.motion3.json")
+            "你好，很高兴见到你。"
+        }
+        handler.speak(greet)
+        handler.appendText("\n[问候] $greet\n")
+        transitionTo(InteractionState.WAITING_REPLY)
+        handler.playMotion("haru_g_m17.motion3.json")
+        scheduleWaitingReplyTimeout()
+    }
+
+    private fun scheduleWaitingReplyTimeout() {
+        waitReplyJob?.cancel()
+        waitReplyJob = scope.launch {
+            hasPendingUserReply = false
+            delay(10_000)
+            if (state != InteractionState.WAITING_REPLY || hasPendingUserReply) return@launch
+            if (facePresent) {
+                enterProactive()
+            } else {
+                enterFarewell()
+            }
+        }
+    }
+
+    private fun enterProactive() {
+        transitionTo(InteractionState.PROACTIVE)
+        proactiveRound = 0
+        askProactiveTopic()
+    }
+
+    private fun askProactiveTopic() {
+        proactiveJob?.cancel()
+        val topics = listOf(
+            "我出一道数学题考考你吧？1+6等于多少？",
+            "我们上完厕所应该干什么呀？",
+            "你喜欢什么颜色呀？",
+        )
+        val idx = proactiveRound.coerceIn(0, topics.lastIndex)
+        val topic = topics[idx]
+        handler.playMotion(if (proactiveRound == 0) "haru_g_m15.motion3.json" else "haru_g_m22.motion3.json")
+        handler.speak(topic)
+        handler.appendText("\n[主动引导] $topic\n")
+        // 将引导内容添加到对话历史中
+        handler.addToChatHistory("助手", topic)
+        // 将引导内容添加到云对话历史中
+        handler.addAssistantMessageToCloudHistory(topic)
+
+        proactiveJob = scope.launch {
+            hasPendingUserReply = false
+            delay(20_000)
+            if (state != InteractionState.PROACTIVE || hasPendingUserReply) return@launch
+            if (!facePresent) {
+                enterFarewell()
+                return@launch
+            }
+            proactiveRound += 1
+            if (proactiveRound < 3) {
+                askProactiveTopic()
+            } else {
+                transitionTo(InteractionState.WAITING_REPLY)
+                handler.playMotion("haru_g_m17.motion3.json")
+                scheduleWaitingReplyTimeout()
+            }
+        }
+    }
+
+    private fun enterFarewell() {
+        transitionTo(InteractionState.FAREWELL)
+        handler.playMotion("haru_g_idle.motion3.json")
+        handler.speak("再见咯")
+        farewellJob?.cancel()
+        farewellJob = scope.launch {
+            delay(3_000)
+            transitionTo(InteractionState.IDLE)
+            handler.playMotion("haru_g_idle.motion3.json")
+            scheduleMemoryMode()
+        }
+    }
+
+    private fun scheduleMemoryMode() {
+        memoryJob?.cancel()
+        if (facePresent) return
+        memoryJob = scope.launch {
+            delay(30_000)
+            if (facePresent || state != InteractionState.IDLE) return@launch
+            transitionTo(InteractionState.MEMORY)
+            handler.playMotion("haru_g_m15.motion3.json")
+            val prompt = "请基于最近互动写一句简短内心想法，口吻温和自然。"
+            handler.requestLocalThought(prompt) { thought ->
+                val finalThought = thought.trim().ifBlank { "我在想，下次见面要聊点有趣的新话题。" }
+                handler.saveThought(finalThought)
+                handler.appendText("\n[回忆] $finalThought\n")
+                if (!facePresent && state == InteractionState.MEMORY) {
+                    transitionTo(InteractionState.IDLE)
+                    handler.playMotion("haru_g_idle.motion3.json")
+                    scheduleMemoryMode()
+                }
+            }
+        }
+    }
+
+    private fun requestDialogueReply(userText: String) {
+        waitReplyJob?.cancel()
+        proactiveJob?.cancel()
+        // 按产品要求：用户对话统一走云端LLM
+        handler.requestCloudReply(userText)
+    }
+
+    private fun transitionTo(newState: InteractionState) {
+        if (state == newState) return
+        state = newState
+        handler.onStateChanged(newState)
+    }
+
+    fun stop() {
+        faceStableJob?.cancel()
+        waitReplyJob?.cancel()
+        proactiveJob?.cancel()
+        memoryJob?.cancel()
+        farewellJob?.cancel()
+    }
+}

app/src/main/java/com/digitalperson/interaction/UserMemoryStore.kt

@@ -0,0 +1,177 @@
+package com.digitalperson.interaction
+
+import android.content.ContentValues
+import android.content.Context
+import android.database.sqlite.SQLiteDatabase
+import android.database.sqlite.SQLiteOpenHelper
+
+data class UserMemory(
+    val userId: String,
+    val displayName: String?,
+    val lastSeenAt: Long,
+    val age: String?,
+    val gender: String?,
+    val hobbies: String?,
+    val preferences: String?,
+    val lastTopics: String?,
+    val lastThought: String?,
+    val profileSummary: String?,
+)
+
+class UserMemoryStore(context: Context) : SQLiteOpenHelper(context, DB_NAME, null, DB_VERSION) {
+    private val memoryCache = LinkedHashMap<String, UserMemory>()
+    @Volatile private var latestThoughtCache: String? = null
+
+    override fun onCreate(db: SQLiteDatabase) {
+        db.execSQL(
+            """
+            CREATE TABLE IF NOT EXISTS user_memory (
+                user_id TEXT PRIMARY KEY,
+                display_name TEXT,
+                last_seen_at INTEGER NOT NULL,
+                age TEXT,
+                gender TEXT,
+                hobbies TEXT,
+                preferences TEXT,
+                last_topics TEXT,
+                last_thought TEXT,
+                profile_summary TEXT
+            )
+            """.trimIndent()
+        )
+    }
+
+    override fun onUpgrade(db: SQLiteDatabase, oldVersion: Int, newVersion: Int) {
+        db.execSQL("DROP TABLE IF EXISTS user_memory")
+        onCreate(db)
+    }
+
+    fun upsertUserSeen(userId: String, displayName: String?) {
+        val existing = memoryCache[userId] ?: getMemory(userId)
+        val now = System.currentTimeMillis()
+        val mergedName = displayName?.takeIf { it.isNotBlank() } ?: existing?.displayName
+        val values = ContentValues().apply {
+            put("user_id", userId)
+            put("display_name", mergedName)
+            put("last_seen_at", now)
+            put("age", existing?.age)
+            put("gender", existing?.gender)
+            put("hobbies", existing?.hobbies)
+            put("preferences", existing?.preferences)
+            put("last_topics", existing?.lastTopics)
+            put("last_thought", existing?.lastThought)
+            put("profile_summary", existing?.profileSummary)
+        }
+        writableDatabase.insertWithOnConflict("user_memory", null, values, SQLiteDatabase.CONFLICT_REPLACE)
+        memoryCache[userId] = UserMemory(
+            userId = userId,
+            displayName = mergedName,
+            lastSeenAt = now,
+            age = existing?.age,
+            gender = existing?.gender,
+            hobbies = existing?.hobbies,
+            preferences = existing?.preferences,
+            lastTopics = existing?.lastTopics,
+            lastThought = existing?.lastThought,
+            profileSummary = existing?.profileSummary,
+        )
+    }
+
+    fun updateDisplayName(userId: String, displayName: String?) {
+        if (displayName.isNullOrBlank()) return
+        upsertUserSeen(userId, displayName)
+    }
+
+    fun updateThought(userId: String, thought: String) {
+        upsertUserSeen(userId, null)
+        val now = System.currentTimeMillis()
+        val values = ContentValues().apply {
+            put("last_thought", thought)
+            put("last_seen_at", now)
+        }
+        writableDatabase.update("user_memory", values, "user_id=?", arrayOf(userId))
+        latestThoughtCache = thought
+        val cached = memoryCache[userId]
+        memoryCache[userId] = UserMemory(
+            userId = userId,
+            displayName = cached?.displayName,
+            lastSeenAt = now,
+            age = cached?.age,
+            gender = cached?.gender,
+            hobbies = cached?.hobbies,
+            preferences = cached?.preferences,
+            lastTopics = cached?.lastTopics,
+            lastThought = thought,
+            profileSummary = cached?.profileSummary,
+        )
+    }
+
+    fun updateProfile(userId: String, age: String?, gender: String?, hobbies: String?, summary: String?) {
+        upsertUserSeen(userId, null)
+        val now = System.currentTimeMillis()
+        val values = ContentValues().apply {
+            if (age != null) put("age", age)
+            if (gender != null) put("gender", gender)
+            if (hobbies != null) put("hobbies", hobbies)
+            if (summary != null) put("profile_summary", summary)
+            put("last_seen_at", now)
+        }
+        writableDatabase.update("user_memory", values, "user_id=?", arrayOf(userId))
+        val cached = memoryCache[userId]
+        memoryCache[userId] = UserMemory(
+            userId = userId,
+            displayName = cached?.displayName,
+            lastSeenAt = now,
+            age = age ?: cached?.age,
+            gender = gender ?: cached?.gender,
+            hobbies = hobbies ?: cached?.hobbies,
+            preferences = cached?.preferences,
+            lastTopics = cached?.lastTopics,
+            lastThought = cached?.lastThought,
+            profileSummary = summary ?: cached?.profileSummary,
+        )
+    }
+
+    fun getMemory(userId: String): UserMemory? {
+        memoryCache[userId]?.let { return it }
+        readableDatabase.rawQuery(
+            "SELECT user_id, display_name, last_seen_at, age, gender, hobbies, preferences, last_topics, last_thought, profile_summary FROM user_memory WHERE user_id=?",
+            arrayOf(userId)
+        ).use { c ->
+            if (!c.moveToFirst()) return null
+            val memory = UserMemory(
+                userId = c.getString(0),
+                displayName = c.getString(1),
+                lastSeenAt = c.getLong(2),
+                age = c.getString(3),
+                gender = c.getString(4),
+                hobbies = c.getString(5),
+                preferences = c.getString(6),
+                lastTopics = c.getString(7),
+                lastThought = c.getString(8),
+                profileSummary = c.getString(9),
+            )
+            memoryCache[userId] = memory
+            if (!memory.lastThought.isNullOrBlank()) {
+                latestThoughtCache = memory.lastThought
+            }
+            return memory
+        }
+    }
+
+    fun getLatestThought(): String? {
+        latestThoughtCache?.let { return it }
+        readableDatabase.rawQuery(
+            "SELECT last_thought FROM user_memory WHERE last_thought IS NOT NULL AND last_thought != '' ORDER BY last_seen_at DESC LIMIT 1",
+            null
+        ).use { c ->
+            if (!c.moveToFirst()) return null
+            return c.getString(0).also { latestThoughtCache = it }
+        }
+    }
+
+    companion object {
+        private const val DB_NAME = "digital_human_memory.db"
+        private const val DB_VERSION = 2
+    }
+}

app/src/main/java/com/digitalperson/ui/Live2DUiManager.kt

@@ -98,19 +98,25 @@ class Live2DUiManager(private val activity: Activity) {
     }
 
     fun appendToUi(s: String) {
-        lastUiText += s
+        activity.runOnUiThread {
-        textView?.text = lastUiText
+            lastUiText += s
-        scrollView?.post { scrollView?.fullScroll(ScrollView.FOCUS_DOWN) }
+            textView?.text = lastUiText
+            scrollView?.post { scrollView?.fullScroll(ScrollView.FOCUS_DOWN) }
+        }
     }
 
     fun clearText() {
-        lastUiText = ""
+        activity.runOnUiThread {
-        textView?.text = ""
+            lastUiText = ""
+            textView?.text = ""
+        }
     }
 
     fun setText(text: String) {
-        lastUiText = text
+        activity.runOnUiThread {
-        textView?.text = text
+            lastUiText = text
+            textView?.text = text
+        }
     }
 
     fun setButtonsEnabled(startEnabled: Boolean = false, stopEnabled: Boolean = false, recordEnabled: Boolean = true) {

app/src/main/java/com/digitalperson/util/FileHelper.kt

@@ -83,13 +83,13 @@ object FileHelper {
 
 //    @JvmStatic
     // 当前使用的模型文件名
-    private const val MODEL_FILE_NAME = "Qwen3-0.6B-rk3588-w8a8.rkllm"
+    private val MODEL_FILE_NAME = com.digitalperson.config.AppConfig.LLM.MODEL_FILE_NAME
     
     fun getLLMModelPath(context: Context): String {
         Log.d(TAG, "=== getLLMModelPath START ===")
 
         // 从应用内部存储目录加载模型
-        val llmDir = ensureDir(File(context.filesDir, "llm"))
+        val llmDir = ensureDir(File(context.filesDir, AppConfig.LLM.MODEL_DIR))
 
         Log.d(TAG, "Loading models from: ${llmDir.absolutePath}")
 
@@ -122,7 +122,7 @@ object FileHelper {
     ) {
         Log.d(TAG, "=== downloadModelFilesWithProgress START ===")
 
-        val llmDir = ensureDir(File(context.filesDir, "llm"))
+        val llmDir = ensureDir(File(context.filesDir, AppConfig.LLM.MODEL_DIR))
 
         // 模型文件列表 - 使用 DeepSeek-R1-Distill-Qwen-1.5B 模型
         val modelFiles = listOf(
@@ -137,19 +137,20 @@ object FileHelper {
                 var totalSize: Long = 0
 
                 // 首先计算总大小
+                val downloadUrl = AppConfig.LLM.DOWNLOAD_SERVER + AppConfig.LLM.DOWNLOAD_PATH
+                Log.i(TAG, "Using download server: ${AppConfig.LLM.DOWNLOAD_SERVER}")
+                
                 for (fileName in modelFiles) {
                     val modelFile = File(llmDir, fileName)
                     if (!modelFile.exists() || modelFile.length() == 0L) {
-                        val size = getFileSizeFromServer("http://192.168.1.19:5000/download/$fileName")
+                        val size = getFileSizeFromServer("$downloadUrl/$fileName")
                         if (size > 0) {
                             totalSize += size
                         } else {
                             // 如果无法获取文件大小，使用估计值
-                            when (fileName) {
+                            val estimatedSize = AppConfig.LLM.MODEL_SIZE_ESTIMATE
-                                MODEL_FILE_NAME -> totalSize += 1L * 1024 * 1024 * 1024 // 1.5B模型约1GB
+                            totalSize += estimatedSize
-                                else -> totalSize += 1L * 1024 * 1024 * 1024 // 1GB 默认
+                            Log.i(TAG, "Using estimated size for $fileName: ${estimatedSize / (1024*1024)} MB")
-                            }
-                            Log.i(TAG, "Using estimated size for $fileName: ${totalSize / (1024*1024)} MB")
                         }
                     }
                 }
@@ -160,7 +161,7 @@ object FileHelper {
                         Log.i(TAG, "Downloading model file: $fileName")
                         try {
                             downloadFileWithProgress(
-                                "http://192.168.1.19:5000/download/$fileName",
+                                "$downloadUrl/$fileName",
                                 modelFile
                             ) { downloaded, total ->
                                 val progress = if (totalSize > 0) {
@@ -199,8 +200,8 @@ object FileHelper {
         return try {
             val connection = java.net.URL(url).openConnection() as java.net.HttpURLConnection
             connection.requestMethod = "HEAD"
-            connection.connectTimeout = 15000
+            connection.connectTimeout = AppConfig.LLM.DOWNLOAD_CONNECT_TIMEOUT
-            connection.readTimeout = 15000
+            connection.readTimeout = AppConfig.LLM.DOWNLOAD_READ_TIMEOUT
             
             // 从响应头获取 Content-Length，避免 int 溢出
             val contentLengthStr = connection.getHeaderField("Content-Length")
@@ -245,8 +246,8 @@ object FileHelper {
         onProgress: (Long, Long) -> Unit
     ) {
         val connection = java.net.URL(url).openConnection() as java.net.HttpURLConnection
-        connection.connectTimeout = 30000
+        connection.connectTimeout = AppConfig.LLM.DOWNLOAD_CONNECT_TIMEOUT
-        connection.readTimeout = 6000000
+        connection.readTimeout = AppConfig.LLM.DOWNLOAD_READ_TIMEOUT
 
         // 从响应头获取 Content-Length，避免 int 溢出
         val contentLengthStr = connection.getHeaderField("Content-Length")
@@ -286,7 +287,7 @@ object FileHelper {
      */
     @JvmStatic
     fun isLocalLLMAvailable(context: Context): Boolean {
-        val llmDir = File(context.filesDir, "llm")
+        val llmDir = File(context.filesDir, AppConfig.LLM.MODEL_DIR)
         
         val rkllmFile = File(llmDir, MODEL_FILE_NAME)