myagent-ai 1.13.3 → 1.13.5

package/memory/manager.py

@@ -633,6 +633,62 @@ class MemoryManager:
     # 记忆搜索
     # ==========================================================================
 
+    # ==========================================================================
+    # 时间衰减权重 (遗忘曲线)
+    # ==========================================================================
+
+    @staticmethod
+    def _compute_time_weight(created_at: str, half_life_days: float = 30.0) -> float:
+        """
+        计算记忆的时间衰减权重（模拟遗忘曲线）。
+
+        使用指数衰减模型:  weight = e^(-ln(2) * age_days / half_life_days)
+        - 刚创建的记忆权重为 1.0
+        - 经过 half_life_days 后权重降至 0.5
+        - 经过 2 * half_life_days 后权重降至 0.25
+        - 权重永远不会降到 0（长期记忆仍可被召回）
+
+        Args:
+            created_at: 记忆创建时间（ISO 8601 格式，如 "2025-01-15T08:30:00+00:00"）
+            half_life_days: 半衰期天数（默认 30 天）
+                - 7 天: 适合短期会话记忆，快速遗忘
+                - 30 天: 默认值，平衡近期与长期记忆
+                - 90 天: 适合知识型场景，长期保留
+
+        Returns:
+            时间权重值 (0.0 ~ 1.0)
+
+        Examples:
+            >>> _compute_time_weight("2025-04-12T00:00:00+00:00", half_life_days=30)
+            1.0  # 今天的记忆
+            >>> _compute_time_weight("2025-03-13T00:00:00+00:00", half_life_days=30)
+            0.5  # 30天前的记忆
+            >>> _compute_time_weight("2025-02-11T00:00:00+00:00", half_life_days=30)
+            0.25  # 60天前的记忆
+        """
+        if not created_at:
+            return 0.5  # 无时间信息的记忆给中等权重
+
+        try:
+            # 解析 ISO 8601 时间戳
+            created_dt = datetime.fromisoformat(created_at.replace("Z", "+00:00"))
+            now_dt = datetime.now(created_dt.tzinfo) if created_dt.tzinfo else datetime.now()
+
+            age_seconds = (now_dt - created_dt).total_seconds()
+            if age_seconds <= 0:
+                return 1.0  # 未来的时间戳或零延迟
+
+            age_days = age_seconds / 86400.0
+
+            # 指数衰减: e^(-ln(2) * age / half_life)
+            decay = math.exp(-0.693147 * age_days / half_life_days)
+
+            # 限制最低权重为 0.05，确保极老的记忆仍有一丝被召回的可能
+            return max(decay, 0.05)
+        except (ValueError, TypeError, OSError) as e:
+            logger.debug(f"时间衰减计算失败 ({created_at}): {e}")
+            return 0.5  # 解析失败时给中等权重
+
     # ==========================================================================
     # TF-IDF 语义搜索 (无外部依赖)
     # ==========================================================================
@@ -753,21 +809,31 @@ class MemoryManager:
         category: str = "",
         limit: int = 10,
         mode: str = "hybrid",
+        time_decay: bool = True,
+        half_life_days: float = 30.0,
     ) -> List[MemoryEntry]:
         """
-        搜索记忆。
+        搜索记忆（内置遗忘曲线时间衰减）。
 
         支持三种搜索模式:
           - "keyword": 传统 LIKE 关键词匹配（快速）
           - "semantic": TF-IDF 语义搜索（理解语义相似性）
           - "hybrid": 混合搜索（默认）= 0.4 * keyword_score + 0.6 * tfidf_score
 
+        所有模式均支持时间衰减权重（遗忘曲线），越久远的记忆权重越低。
+        最终得分 = 相关性得分 × 时间权重。
+
         Args:
             query: 搜索查询
             session_id: 会话 ID（空=所有会话）
             category: 记忆类别（空=所有类别）
             limit: 返回数量
             mode: 搜索模式 "keyword" | "semantic" | "hybrid"
+            time_decay: 是否启用时间衰减（默认 True）
+            half_life_days: 遗忘曲线半衰期天数（默认 30 天）
+                - 7 天: 短期记忆场景，快速遗忘
+                - 30 天: 默认，平衡近期与长期
+                - 90 天: 知识型场景，长期保留
         """
         conn = self._get_conn()
         conditions = ["1=1"]
@@ -787,26 +853,38 @@ class MemoryManager:
         where = " AND ".join(conditions)
 
         if mode == "keyword":
-            return self._search_keyword(conn, query, where, params, limit)
+            return self._search_keyword(conn, query, where, params, limit,
+                                         time_decay=time_decay, half_life_days=half_life_days)
         elif mode == "semantic":
-            return self._search_semantic(conn, query, where, params, limit)
+            return self._search_semantic(conn, query, where, params, limit,
+                                         time_decay=time_decay, half_life_days=half_life_days)
         else:
-            # 混合模式：取两种搜索结果的加权和
-            keyword_results = self._search_keyword(conn, query, where, params, limit * 2)
-            semantic_results = self._search_semantic(conn, query, where, params, limit * 2)
+            # 混合模式：取两种搜索结果的加权和 + 时间衰减
+            keyword_results = self._search_keyword(conn, query, where, params, limit * 2,
+                                                   time_decay=False, half_life_days=half_life_days)
+            semantic_results = self._search_semantic(conn, query, where, params, limit * 2,
+                                                     time_decay=False, half_life_days=half_life_days)
 
-            # 合并评分
+            # 合并评分（相关性 + 时间衰减）
             combined: Dict[str, Tuple[MemoryEntry, float]] = {}
             for i, entry in enumerate(keyword_results):
-                score = 1.0 - (i / max(len(keyword_results), 1))
-                combined[entry.id] = (entry, score * 0.4)
+                relevance = 1.0 - (i / max(len(keyword_results), 1))
+                score = relevance * 0.4
+                combined[entry.id] = (entry, score)
 
             for i, entry in enumerate(semantic_results):
-                score = 1.0 - (i / max(len(semantic_results), 1))
+                relevance = 1.0 - (i / max(len(semantic_results), 1))
+                sem_score = relevance * 0.6
                 if entry.id in combined:
-                    combined[entry.id] = (entry, combined[entry.id][1] + score * 0.6)
+                    combined[entry.id] = (entry, combined[entry.id][1] + sem_score)
                 else:
-                    combined[entry.id] = (entry, score * 0.6)
+                    combined[entry.id] = (entry, sem_score)
+
+            # 应用时间衰减权重: final_score = relevance_score × time_weight
+            if time_decay:
+                for mem_id, (entry, relevance_score) in combined.items():
+                    tw = self._compute_time_weight(entry.created_at, half_life_days)
+                    combined[mem_id] = (entry, relevance_score * tw)
 
             # 按综合得分排序
             sorted_results = sorted(
@@ -832,28 +910,45 @@ class MemoryManager:
         where: str,
         params: list,
         limit: int,
+        time_decay: bool = False,
+        half_life_days: float = 30.0,
     ) -> List[MemoryEntry]:
-        """关键词 LIKE 搜索"""
+        """关键词 LIKE 搜索（支持时间衰减重排序）"""
         like_pattern = f"%{query}%"
         conditions = f"{where} AND (content LIKE ? OR summary LIKE ? OR key LIKE ?)"
         search_params = params + [like_pattern, like_pattern, like_pattern]
 
+        # 取更多候选（后续按时间衰减重排序）
+        fetch_limit = limit * 3 if time_decay else limit
         sql = f"""
             SELECT * FROM memories WHERE {conditions}
             ORDER BY importance DESC, access_count DESC
             LIMIT ?
         """
-        search_params.append(limit)
+        search_params.append(fetch_limit)
         rows = conn.execute(sql, search_params).fetchall()
 
-        for row in rows:
+        entries = [MemoryEntry.from_row(row) for row in rows]
+
+        # 应用时间衰减重排序
+        if time_decay and entries:
+            scored = []
+            for entry in entries:
+                base_score = entry.importance + entry.access_count * 0.01
+                tw = self._compute_time_weight(entry.created_at, half_life_days)
+                scored.append((entry, base_score * tw))
+            scored.sort(key=lambda x: x[1], reverse=True)
+            entries = [e for e, _ in scored[:limit]]
+
+        # 更新访问计数
+        for entry in entries:
             conn.execute(
                 "UPDATE memories SET access_count = access_count + 1 WHERE id = ?",
-                (row["id"],),
+                (entry.id,),
             )
         conn.commit()
 
-        return [MemoryEntry.from_row(row) for row in rows]
+        return entries
 
     def _search_semantic(
         self,
@@ -862,8 +957,10 @@ class MemoryManager:
         where: str,
         params: list,
         limit: int,
+        time_decay: bool = False,
+        half_life_days: float = 30.0,
     ) -> List[MemoryEntry]:
-        """TF-IDF 语义搜索"""
+        """TF-IDF 语义搜索（支持时间衰减重排序）"""
         # 先取一批候选文档
         candidate_sql = f"SELECT * FROM memories WHERE {where} ORDER BY created_at DESC LIMIT 200"
         rows = conn.execute(candidate_sql, params).fetchall()
@@ -883,6 +980,14 @@ class MemoryManager:
         # 计算 TF-IDF 得分
         scores = self._compute_tfidf(query, documents)
 
+        # 应用时间衰减重排序
+        if time_decay:
+            for doc_id, tfidf_score in scores.items():
+                row = row_map.get(doc_id)
+                if row:
+                    tw = self._compute_time_weight(row["created_at"], half_life_days)
+                    scores[doc_id] = tfidf_score * tw
+
         # 按得分排序，取 top N
         top_ids = list(scores.keys())[:limit]
         result = [MemoryEntry.from_row(row_map[doc_id]) for doc_id in top_ids if doc_id in row_map]
@@ -903,9 +1008,12 @@ class MemoryManager:
         category: str = "",
         limit: int = 20,
         mode: str = "hybrid",
+        time_decay: bool = True,
+        half_life_days: float = 30.0,
     ) -> List[MemoryEntry]:
         """跨会话搜索"""
-        return self.search(query, session_id="", category=category, limit=limit, mode=mode)
+        return self.search(query, session_id="", category=category, limit=limit, mode=mode,
+                          time_decay=time_decay, half_life_days=half_life_days)
 
     # ==========================================================================
     # 记忆总结与维护

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "myagent-ai",
-  "version": "1.13.3",
+  "version": "1.13.5",
   "description": "本地桌面端执行型AI助手 - Open Interpreter 风格 | Local Desktop Execution-Oriented AI Assistant",
   "main": "main.py",
   "bin": {

package/web/ui/chat/chat_main.js

@@ -2419,14 +2419,14 @@ function _renderMessagesInner() {
         partsInner += '<div class="timeline-segment">' + renderMarkdown(msg._streamingText) + _cursor + '</div>';
       }
       if (partsInner) {
-        // All parts (text segments + tool calls) wrapped in ONE message-bubble
-        timelineHtml = '<div class="message-bubble"><div class="msg-timeline">' + partsInner + '</div></div>';
+        // All parts (text segments + tool calls) wrapped in ONE message-bubble (full width during streaming)
+        timelineHtml = '<div class="message-bubble msg-bubble-wrapper"><div class="msg-timeline">' + partsInner + '</div></div>';
       }
     }
 
-    // Backward compat: single bubble for messages without parts
+    // Backward compat: single bubble for messages without parts (full width)
     const singleBubbleHtml = (!hasParts && !hasStreamingText)
-      ? (content ? `<div class="message-bubble">${content}${ttsIndicator}</div>` : '')
+      ? (content ? `<div class="message-bubble msg-bubble-wrapper">${content}${ttsIndicator}</div>` : '')
       : '';
 
     // Exec events panel: only for backward compat (messages without parts loaded from DB)
@@ -3459,6 +3459,9 @@ const ttsManager = {
   _audioPlaying: false,     // 队列是否正在播放
   _stopRequested: false,    // 是否已请求停止
   _streamMsgIndex: -1,      // 流式模式对应的消息索引
+  _pendingFetches: 0,       // 正在进行的 TTS 请求计数（防竞态）
+  _flushed: false,          // streamFlush 已调用（防竞态停止）
+  _firstChunkSent: false,   // 是否已发送第一个音频块（降低首句延迟）
 
   init() {
     // Load TTS enabled state from localStorage
@@ -3545,6 +3548,9 @@ const ttsManager = {
     this._audioQueue = [];
     this._audioPlaying = false;
     this._stopRequested = false;
+    this._pendingFetches = 0;
+    this._flushed = false;
+    this._firstChunkSent = false;
     // 保持 _streamActive = true 和 isPlaying = true
     // 这样后续的 text_delta 会继续往新的缓冲区添加文本
     // 新的句子会被合成并入队播放
@@ -3574,6 +3580,9 @@ const ttsManager = {
     this._streamMsgIndex = msgIndex;
     this.currentMsgIndex = msgIndex;
     this.isPlaying = true;
+    this._pendingFetches = 0;
+    this._flushed = false;
+    this._firstChunkSent = false;
   },
 
   /**
@@ -3588,6 +3597,27 @@ const ttsManager = {
 
     this._streamBuffer += delta;
 
+    // 首句快速通道：第一个音频块降低延迟，缓冲区达到 25 字即触发
+    if (!this._firstChunkSent && this._streamBuffer.length >= 25) {
+      var lastSep = -1;
+      for (var k = 0; k < this._streamBuffer.length; k++) {
+        var c = this._streamBuffer[k];
+        if (c === '，' || c === ',' || c === '。' || c === '！' || c === '？' || c === '；' || c === ' ' || c === '\n') {
+          lastSep = k;
+        }
+      }
+      if (lastSep > 0) {
+        var firstChunk = this._streamBuffer.substring(0, lastSep + 1).trim();
+        this._streamBuffer = this._streamBuffer.substring(lastSep + 1);
+        var cleanFirst = this._cleanForStreamTTS(firstChunk);
+        if (cleanFirst) {
+          this._firstChunkSent = true;
+          this._enqueueTTS(cleanFirst);
+        }
+      }
+      return;
+    }
+
     // 检测句子边界：中文句号/感叹号/问号，英文句号+空格，或换行
     var boundaryPattern = /[。！？]|\.(?:\s|$)|\n/;
     var boundaryIdx = -1;
@@ -3643,8 +3673,15 @@ const ttsManager = {
         this._enqueueTTS(cleanText);
       }
     }
-    // 标记流式阶段结束（队列播完后自动清理状态）
-    this._streamActive = false;
+    // 标记已刷新，但不立即停止 —— 等队列和 pending fetch 全部完成后再停止
+    this._flushed = true;
+    // 如果没有 pending fetch 且队列为空，立即停止
+    if (this._pendingFetches === 0 && this._audioQueue.length === 0 && !this._audioPlaying) {
+      this.isPlaying = false;
+      this._streamActive = false;
+      this.currentMsgIndex = -1;
+      this.updatePlayingIndicator();
+    }
   },
 
   /**
@@ -3683,6 +3720,7 @@ const ttsManager = {
   _enqueueTTS(text) {
     if (this._stopRequested) return;
     var self = this;
+    self._pendingFetches++;
 
     (async function() {
       try {
@@ -3719,6 +3757,15 @@ const ttsManager = {
         }
       } catch (e) {
         console.error('TTS stream chunk error:', e);
+      } finally {
+        self._pendingFetches--;
+        // 如果已 flush 且没有更多 pending，检查是否该停止
+        if (self._flushed && self._pendingFetches === 0 && self._audioQueue.length === 0 && !self._audioPlaying) {
+          self.isPlaying = false;
+          self._streamActive = false;
+          self.currentMsgIndex = -1;
+          self.updatePlayingIndicator();
+        }
       }
     })();
   },
@@ -3736,15 +3783,16 @@ const ttsManager = {
     }
 
     if (this._audioQueue.length === 0) {
-      // 队列空了，检查流式是否已结束
-      if (!this._streamActive) {
-        // 流结束且队列为空 → 播放完成
+      // 队列空了，检查流式是否已结束 AND 没有 pending fetches
+      if (this._flushed && this._pendingFetches === 0) {
+        // 流已结束、无 pending 请求、队列空 → 播放完成
         this.isPlaying = false;
+        this._streamActive = false;
         this._audioPlaying = false;
         this.currentMsgIndex = -1;
         this.updatePlayingIndicator();
       }
-      // 如果流还在继续，等待新的音频入队
+      // 如果流还在继续或还有 pending fetches，等待新的音频入队
       return;
     }
 

package/web/ui/chat/flow_engine.js

@@ -472,59 +472,7 @@ function updateStreamingMessage(msgIdx) {
     }
   }
 
-  // V2 Reasoning block (from v2_reasoning events)
-  if (msg._v2Reasoning) {
-    let v2ReasoningBlock = null;
-    const allThoughts2 = contentArea.querySelectorAll('.thought-block');
-    for (const tb of allThoughts2) {
-      const label = tb.querySelector('.thought-label');
-      if (label && label.textContent.includes('V2 推理')) {
-        v2ReasoningBlock = tb;
-        break;
-      }
-    }
-    if (!v2ReasoningBlock && msg.thought) {
-      // Skip if V1 thought already exists - don't duplicate
-    } else {
-      const v2Len = msg._v2Reasoning.length;
-      const v2WordCount = msg.streaming
-        ? '<span class="thought-word-count">' + v2Len + ' 字</span>'
-        : '';
-      if (v2ReasoningBlock) {
-        // Incremental update for V2 reasoning block
-        const label = v2ReasoningBlock.querySelector('.thought-label');
-        if (label) label.innerHTML = 'V2 推理过程' + v2WordCount;
-        const badge = v2ReasoningBlock.querySelector('.thought-badge');
-        if (badge) badge.textContent = msg.streaming ? '推理中...' : '已完成';
-        const thoughtContent = v2ReasoningBlock.querySelector('.thought-content');
-        if (thoughtContent && msg.streaming) {
-          const prevLen = v2ReasoningBlock._lastV2Len || 0;
-          if (msg._v2Reasoning.length > prevLen) {
-            const newText = msg._v2Reasoning.substring(prevLen);
-            thoughtContent.insertAdjacentHTML('beforeend', renderMarkdown(newText));
-            v2ReasoningBlock._lastV2Len = msg._v2Reasoning.length;
-            // 自动滚动 V2 推理框内部内容到底部
-            thoughtContent.scrollTop = thoughtContent.scrollHeight;
-          }
-        } else if (thoughtContent && !msg.streaming) {
-          thoughtContent.innerHTML = renderMarkdown(msg._v2Reasoning);
-          v2ReasoningBlock._lastV2Len = msg._v2Reasoning.length;
-        }
-      } else if (!msg.thought) {
-        const v2Html = `<details class="thought-block ${msg.streaming ? 'streaming' : ''}" ${msg.streaming ? 'open' : ''}>
-          <summary>
-            <span class="thought-icon">🧠</span>
-            <span class="thought-label">V2 推理过程${v2WordCount}</span>
-            ${msg.streaming ? '<span class="thought-badge">推理中...</span>' : '<span class="thought-badge">已完成</span>'}
-          </summary>
-          <div class="thought-content">${renderMarkdown(msg._v2Reasoning)}</div>
-        </details>`;
-        contentArea.insertAdjacentHTML('afterbegin', v2Html);
-        const newBlock = contentArea.querySelector(':scope > .thought-block');
-        if (newBlock) newBlock._lastV2Len = msg._v2Reasoning.length;
-      }
-    }
-  }
+  // V2 Reasoning: content is already rendered in msg.parts timeline, skip separate block to avoid duplication
 
   // Update content - timeline (interleaved text + exec events) or single bubble (backward compat)
   const hasParts = Array.isArray(msg.parts);