myagent-ai 1.9.4 → 1.9.6

package/agents/main_agent.py

@@ -188,7 +188,7 @@ status 取值：
 
 <tool><beforecalltext>连接词，介绍调用什么工具，达到什么目的。示例:接下来，要调用命令行工具，获得ip地址。</beforecalltext><toolname>工具名</toolname><parms>参数</parms><timeout>预估超时时限</timeout><callback>true/false</callback></tool>
 </toolstocal>
-<remember>根据"context,提取需要存入记忆库的内容，这里要搜简上记忆库</remember>
+<remember>仅从最新用户输入（userprint 或 usersays_correct）中提炼值得长期记忆的信息（如用户偏好、重要结论、错误经验等）。不要从历史对话中重复提炼旧记忆。如果本轮用户输入没有新信息需要记忆，则为空。</remember>
 <recall>下一轮执行需要调取的记忆，这里要设计接上记忆库</recall>
 <get_knowledge>下一轮执行时需要从知识库搜索获得的知识，填写检索关键词或描述。如context中已包含充足的<knowledge>内容，则为空。如需更多专业知识支撑，则填写相关搜索词。</get_knowledge>
 <askuser>需要询问用户的内容，如无，则为空</askuser>
@@ -203,7 +203,7 @@ status 取值：
 4. <toolstocal>: 列出所有需要执行的工具调用，每个工具包含完整的参数说明
 5. <timeout>: 预估超时秒数（简单操作10-30s，文件操作30-60s，网络请求60-120s，数据处理120-300s）
 6. <callback>: 如果该工具的执行结果对后续决策有影响，设为 true；否则设为 false
-7. <remember>: 提取本轮对话中值得长期记忆的关键信息（用户偏好、重要结论、错误经验等）
+7. <remember>: 仅从最新用户输入（userprint 或 usersays_correct）中提炼值得长期记忆的关键信息，不要重复提炼历史对话中已有的记忆。如果本轮没有新信息需要记忆，则为空
 8. <recall>: 描述下一轮执行时需要从记忆库中检索的内容关键词
 9. <get_knowledge>: 如果当前 <knowledge> 内容不足以完成任务，填写需要从知识库搜索的关键词；否则为空
 10. <askuser>: 当信息不足需要用户补充时，在此填写要问的问题
@@ -1139,6 +1139,81 @@ status 取值：
         else:
             logger.debug(f"[v2-event] {event_type}: {data}")
 
+    async def _merge_duplicate_memory(
+        self,
+        old_memory,
+        new_content: str,
+        context: AgentContext,
+        task_id: str,
+    ) -> Optional[str]:
+        """
+        当发现新旧记忆高度相似时，调用 LLM API 让其判断最终记忆内容。
+
+        将旧记忆、新记忆、当前上下文发送给 LLM，由 LLM 决定:
+          - 合并为一条更完整的记忆
+          - 保留新记忆（旧记忆已过时）
+          - 保留旧记忆（新记忆无新增信息）
+
+        Returns:
+            合并后的记忆内容，或 None（合并失败）
+        """
+        if not self.llm:
+            logger.warning(f"[{task_id}] 记忆合并: 无 LLM 客户端，跳过合并")
+            return None
+
+        from datetime import datetime
+        old_time = old_memory.created_at or "未知时间"
+        new_time = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
+        user_msg = context.user_message or ""
+
+        merge_prompt = f"""你是一个记忆管理系统。现在系统检测到两条高度相似的记忆，请你判断如何合并它们。
+
+## 旧记忆（创建于 {old_time}）
+{old_memory.content}
+
+## 新记忆（创建于 {new_time}）
+{new_content}
+
+## 当前用户输入
+{user_msg}
+
+## 任务
+请分析新旧记忆，输出一条最终的合并记忆。规则：
+1. 如果新记忆包含了旧记忆的信息并有更新，则合并为更完整的表述
+2. 如果新记忆只是旧记忆的重复或信息量更少，保留旧记忆中更完整的信息
+3. 如果新记忆提供了全新的信息，以新记忆为主，补充旧记忆中的有效部分
+4. 合并后的记忆应当简洁、准确、包含时间上下文
+5. 直接输出合并后的记忆内容，不要输出任何解释或标记
+
+请输出合并后的记忆："""
+
+        try:
+            messages = [
+                Message(role="system", content="你是一个记忆管理系统，负责合并重复或相似的记忆条目。只输出合并后的记忆内容，不要输出任何额外说明。"),
+                Message(role="user", content=merge_prompt),
+            ]
+
+            response = await self._call_llm(messages)
+
+            if response.success and response.content:
+                merged = response.content.strip()
+                # 清理可能的引号包裹
+                if merged.startswith('"') and merged.endswith('"'):
+                    merged = merged[1:-1]
+                if merged.startswith("'") and merged.endswith("'"):
+                    merged = merged[1:-1]
+                logger.info(
+                    f"[{task_id}] 记忆合并成功: 旧({len(old_memory.content)}字) + "
+                    f"新({len(new_content)}字) → 合并({len(merged)}字)"
+                )
+                return merged
+            else:
+                logger.warning(f"[{task_id}] 记忆合并 LLM 调用失败: {response.error}")
+                return None
+        except Exception as e:
+            logger.warning(f"[{task_id}] 记忆合并异常: {e}")
+            return None
+
     async def process_v2(
         self,
         context: AgentContext,
@@ -1210,6 +1285,8 @@ status 取值：
         all_tool_outputs = ""
         recall_content = ""
         get_knowledge_content = ""
+        # 追踪流式推送的 reasoning 文本（用于构建有意义的最终回复）
+        _v2_reasoning_collected: List[str] = []
 
         conversation_history = list(context.conversation_history or [])
 
@@ -1347,39 +1424,68 @@ status 取值：
             if parsed.usersays_correct:
                 context.working_memory["usersays_correct"] = parsed.usersays_correct
 
-            # Step 6: 处理 remember — 存入长期记忆
-            if parsed.remember and self.memory_agent:
+            # Step 6: 处理 remember — 查重+LLM合并后存入长期记忆
+            if parsed.remember:
                 try:
-                    mem_ctx = AgentContext(
-                        task_id=task_id,
-                        session_id=context.session_id,
-                        metadata={
-                            "memory_action": "save",
-                            "content": parsed.remember,
-                        },
-                    )
-                    await self.memory_agent.process(mem_ctx)
-                    await self._emit_v2_event(
-                        "v2_memory_saved",
-                        {"content": truncate_str(parsed.remember, 200)},
-                        stream_callback,
-                    )
-                except Exception as e:
-                    logger.warning(f"[{task_id}] 存入记忆失败: {e}")
-            elif parsed.remember and self.memory:
-                try:
-                    self.memory.add_long_term(
-                        session_id=context.session_id,
-                        key="conversation_insight",
-                        content=parsed.remember,
-                        summary=truncate_str(parsed.remember, 200),
-                        importance=0.7,
-                    )
-                    await self._emit_v2_event(
-                        "v2_memory_saved",
-                        {"content": truncate_str(parsed.remember, 200)},
-                        stream_callback,
-                    )
+                    if self.memory:
+                        # 查找是否有相似记忆
+                        dup_memory = self.memory.find_duplicate_memory(
+                            content=parsed.remember,
+                            session_id=context.session_id,
+                            key="conversation_insight",
+                        )
+                        if dup_memory:
+                            # 发现相似记忆 → 调用 LLM API 合并新旧记忆
+                            logger.info(
+                                f"[{task_id}] 记忆查重: 发现相似内容，调用LLM合并 "
+                                f"(旧记忆ID={dup_memory.id}, 创建于={dup_memory.created_at})"
+                            )
+                            merged_content = await self._merge_duplicate_memory(
+                                old_memory=dup_memory,
+                                new_content=parsed.remember,
+                                context=context,
+                                task_id=task_id,
+                            )
+                            if merged_content:
+                                # 用 LLM 合并后的内容替换旧记忆
+                                self.memory.update_memory(
+                                    memory_id=dup_memory.id,
+                                    content=merged_content,
+                                    summary=truncate_str(merged_content, 200),
+                                )
+                                logger.info(f"[{task_id}] 记忆已合并更新: {dup_memory.id}")
+                            else:
+                                # LLM 合并失败，直接更新为新内容
+                                self.memory.update_memory(
+                                    memory_id=dup_memory.id,
+                                    content=parsed.remember,
+                                )
+                                logger.info(f"[{task_id}] 记忆直接更新为新内容: {dup_memory.id}")
+                        else:
+                            # 无重复，直接存储新记忆
+                            if self.memory_agent:
+                                mem_ctx = AgentContext(
+                                    task_id=task_id,
+                                    session_id=context.session_id,
+                                    metadata={
+                                        "memory_action": "save",
+                                        "content": parsed.remember,
+                                    },
+                                )
+                                await self.memory_agent.process(mem_ctx)
+                            else:
+                                self.memory.add_long_term(
+                                    session_id=context.session_id,
+                                    key="conversation_insight",
+                                    content=parsed.remember,
+                                    summary=truncate_str(parsed.remember, 200),
+                                    importance=0.7,
+                                )
+                        await self._emit_v2_event(
+                            "v2_memory_saved",
+                            {"content": truncate_str(parsed.remember, 200)},
+                            stream_callback,
+                        )
                 except Exception as e:
                     logger.warning(f"[{task_id}] 存入记忆失败: {e}")
 
@@ -1470,6 +1576,7 @@ status 取值：
 
                 # 发送 beforecalltext 作为显示文本
                 if before_call:
+                    _v2_reasoning_collected.append(before_call)
                     await self._emit_v2_event(
                         "v2_reasoning",
                         {"content": before_call},
@@ -1548,9 +1655,15 @@ status 取值：
             # 核心逻辑: finish=true 表示任务已完成/不需要再调用LLM，即使工具设置了callback=true
             if parsed.finish:
                 logger.info(f"[{task_id}] finish=true，任务已完成，不回调 LLM")
-                final_text = "已完成所有操作。"
-                if current_task_plan:
-                    final_text += f"\n\n任务计划:\n{current_task_plan}"
+                # 构建有意义的最终回复：使用收集到的 reasoning text + 任务计划摘要
+                if _v2_reasoning_collected:
+                    final_text = "\n".join(_v2_reasoning_collected)
+                    if current_task_plan:
+                        final_text += f"\n\n{current_task_plan}"
+                else:
+                    final_text = "已完成所有操作。"
+                    if current_task_plan:
+                        final_text += f"\n\n任务计划:\n{current_task_plan}"
                 context.working_memory["final_response"] = final_text
                 await self._emit_v2_event("v2_reasoning", {"content": truncate_str(final_text, 3000)}, stream_callback)
                 if self.memory:
@@ -1564,9 +1677,14 @@ status 取值：
             # finish=false: 根据工具的 callback 标志决定是否回调
             if not need_callback:
                 logger.info(f"[{task_id}] 所有工具无需回调且 finish=false，结束循环")
-                final_text = "已完成所有操作。"
-                if current_task_plan:
-                    final_text += f"\n\n任务计划:\n{current_task_plan}"
+                if _v2_reasoning_collected:
+                    final_text = "\n".join(_v2_reasoning_collected)
+                    if current_task_plan:
+                        final_text += f"\n\n{current_task_plan}"
+                else:
+                    final_text = "已完成所有操作。"
+                    if current_task_plan:
+                        final_text += f"\n\n任务计划:\n{current_task_plan}"
                 context.working_memory["final_response"] = final_text
                 await self._emit_v2_event("v2_reasoning", {"content": truncate_str(final_text, 3000)}, stream_callback)
                 if self.memory:

package/core/output_parser.py

@@ -20,7 +20,7 @@ Expected XML schema produced by the LLM::
                 <callback>true/false</callback>
             </tool>
         </toolstocal>
-        <remember>需要存入记忆库的内容</remember>
+        <remember>仅从最新用户输入中提炼的记忆，无新信息则为空</remember>
         <recall>下一轮需要调取的记忆</recall>
         <askuser>需要询问用户的内容</askuser>
         <get_knowledge>下一轮需要搜索获得的知识</get_knowledge>

package/main.py

@@ -319,6 +319,11 @@ class MyAgentApp:
             memory_agent=self.memory_agent,
             config_broadcaster=self.config_broadcaster,
         )
+        self.main_agent.init_context_builder(
+            memory_manager=self.memory,
+            skill_registry=self.skill_registry,
+            knowledge_base_dir=str(self.config_mgr.data_dir / "organization" / "knowledge")
+        )
         self.logger.info("Agent 集群已初始化 (主/工具/记忆)")
 
         # 10. 自动更新管理器

package/memory/manager.py

@@ -479,6 +479,128 @@ class MemoryManager:
     # 长期记忆 (持久知识)
     # ==========================================================================
 
+    def find_duplicate_memory(
+        self,
+        content: str,
+        session_id: str = "",
+        key: str = "",
+        similarity_threshold: float = 0.85,
+    ) -> Optional[MemoryEntry]:
+        """
+        查找与给定内容高度相似的已有记忆（查重）。
+
+        使用 TF-IDF 余弦相似度判断新内容是否与已有记忆高度重复。
+        相似度超过阈值则返回匹配的旧记忆，否则返回 None。
+
+        Args:
+            content: 待检查的新记忆内容
+            session_id: 会话 ID（为空则跨会话查重）
+            key: 记忆分类键（为空则不限制分类）
+            similarity_threshold: 相似度阈值，超过此值视为重复（默认 0.85）
+
+        Returns:
+            匹配的旧 MemoryEntry，或 None（无重复）
+        """
+        if not content or not content.strip():
+            return None  # 空内容不查重
+
+        conn = self._get_conn()
+        conditions = ["category = 'long_term'"]
+        params: list = []
+
+        if session_id:
+            conditions.append("session_id = ?")
+            params.append(session_id)
+        if key:
+            conditions.append("key = ?")
+            params.append(key)
+
+        # 过滤过期
+        conditions.append("(expires_at = '' OR expires_at > ?)")
+        params.append(timestamp())
+
+        where = " AND ".join(conditions)
+        # 取最近的记忆候选进行比对
+        sql = f"SELECT id, content, summary, created_at FROM memories WHERE {where} ORDER BY created_at DESC LIMIT 50"
+        rows = conn.execute(sql, params).fetchall()
+
+        if not rows:
+            return None
+
+        # 使用 TF-IDF 计算新内容与每条已有记忆的相似度
+        documents = [(row["id"], row["content"]) for row in rows]
+        scores = self._compute_tfidf(content, documents)
+
+        if not scores:
+            return None
+
+        max_id = max(scores, key=scores.get)
+        max_score = scores[max_id]
+        if max_score >= similarity_threshold:
+            logger.debug(
+                f"记忆查重: 发现相似内容 (相似度={max_score:.2f}, "
+                f"阈值={similarity_threshold}, 匹配ID={max_id})"
+            )
+            # 返回匹配的旧记忆
+            for row in rows:
+                if row["id"] == max_id:
+                    return MemoryEntry.from_row(row)
+
+        return None
+
+    def is_duplicate_memory(
+        self,
+        content: str,
+        session_id: str = "",
+        key: str = "",
+        similarity_threshold: float = 0.85,
+    ) -> bool:
+        """
+        检查是否已存在高度相似的记忆（查重）。
+
+        Returns:
+            True 表示存在重复记忆，False 表示可以存储
+        """
+        return self.find_duplicate_memory(
+            content, session_id, key, similarity_threshold
+        ) is not None
+
+    def update_memory(self, memory_id: str, content: str, summary: str = "", **updates) -> bool:
+        """
+        更新记忆内容（保留原 ID 和创建时间）。
+
+        Args:
+            memory_id: 记忆 ID
+            content: 新内容
+            summary: 新摘要（为空则自动截取）
+
+        Returns:
+            True 更新成功
+        """
+        if not summary:
+            summary = truncate_str(content, 200)
+        self._update_content(memory_id, content, summary=summary, **updates)
+        logger.info(f"记忆已更新: {memory_id}")
+        return True
+
+    def delete_memory(self, memory_id: str) -> bool:
+        """
+        删除单条记忆。
+
+        Args:
+            memory_id: 记忆 ID
+
+        Returns:
+            True 删除成功
+        """
+        conn = self._get_conn()
+        cursor = conn.execute("DELETE FROM memories WHERE id = ?", (memory_id,))
+        conn.commit()
+        if cursor.rowcount > 0:
+            logger.info(f"记忆已删除: {memory_id}")
+            return True
+        return False
+
     def add_long_term(
         self,
         session_id: str = "global",
@@ -498,14 +620,23 @@ class MemoryManager:
             summary: 简要摘要
             importance: 重要性(越高越不容易被淘汰)
         """
+        from datetime import datetime
+        now_str = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
+        # 在 content 开头自动附加时间戳
+        timestamped_content = f"[{now_str}] {truncate_str(content, 50000)}"
+        ts_summary = summary or truncate_str(content, 200)
+
         entry = MemoryEntry(
             session_id=session_id,
             category="long_term",
             key=key,
-            content=truncate_str(content, 50000),
-            summary=summary,
+            content=timestamped_content,
+            summary=f"[{now_str}] {ts_summary}",
             importance=importance,
-            metadata=metadata or {},
+            metadata={
+                "timestamp": now_str,
+                **(metadata or {}),
+            },
         )
         return self._insert(entry)
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "myagent-ai",
-  "version": "1.9.4",
+  "version": "1.9.6",
   "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/api_server.py

@@ -3063,8 +3063,19 @@ class ApiServer:
 
             # 返回最终响应
             final_response = v2_context.working_memory.get("final_response", "")
-            if not final_response and _v2_reasoning_parts:
-                final_response = "\n".join(_v2_reasoning_parts)
+            # 优先使用 v2_reasoning 流式推送的文本（用户实际看到的内容），
+            # 而不是后端设定的通用占位符如"已完成所有操作。"
+            if _v2_reasoning_parts:
+                # Smart join: only add separator when parts don't already end with newline
+                joined_parts = []
+                for i, part in enumerate(_v2_reasoning_parts):
+                    if i > 0 and joined_parts and not joined_parts[-1].endswith('\n') and not part.startswith('\n'):
+                        joined_parts.append('\n')
+                    joined_parts.append(part)
+                reasoning_text = ''.join(joined_parts)
+                # 如果 final_response 是通用占位符或为空，使用 reasoning text
+                if not final_response or final_response.startswith("已完成所有操作") or final_response.startswith("处理完毕"):
+                    final_response = reasoning_text
             return final_response
 
         # ── V1 路由: 标准 JSON action 格式 ──

package/web/ui/chat/flow_engine.js

@@ -814,6 +814,29 @@ function showToolResultModal(msgIndex, eventId) {
   document.body.appendChild(overlay);
 }
 
+// ══════════════════════════════════════════════════════
+// ── V2 Helpers ──
+// ══════════════════════════════════════════════════════
+
+/**
+ * Strip XML tags from text for real-time streaming preview in V2 mode.
+ * Shows plain text between tags so the user sees progress during LLM streaming.
+ */
+function _stripXmlTags(xml) {
+  if (!xml) return '';
+  // Remove all XML tags, collapse excessive whitespace
+  return xml
+    .replace(/<[^>]+>/g, ' ')     // Replace tags with space
+    .replace(/&lt;/g, '<')
+    .replace(/&gt;/g, '>')
+    .replace(/&amp;/g, '&')
+    .replace(/&quot;/g, '"')
+    .replace(/&#39;/g, "'")
+    .replace(/&#x27;/g, "'")
+    .replace(/\s{3,}/g, ' ')      // Collapse 3+ whitespace to single space
+    .trim();
+}
+
 // ══════════════════════════════════════════════════════
 // ── V2 Content Assembler (V2 内容组装) ──
 // ══════════════════════════════════════════════════════
@@ -828,9 +851,12 @@ function _assembleV2Content(msg, msgParts) {
     return msg._askUser.trim();
   }
   // Priority 3: V1 text parts (backward compat — non-V2 mode)
-  var textParts = msgParts.filter(function(p) { return p.type === 'text'; });
-  if (textParts.length > 0) {
-    return textParts.map(function(p) { return p.content; }).join('\n\n');
+  // Guard: msgParts may be undefined after page refresh (only role/content/time persisted)
+  if (msgParts && Array.isArray(msgParts) && msgParts.length > 0) {
+    var textParts = msgParts.filter(function(p) { return p.type === 'text'; });
+    if (textParts.length > 0) {
+      return textParts.map(function(p) { return p.content; }).join('\n\n');
+    }
   }
   // Priority 4: raw content from message (server-stored response)
   if (msg.content && msg.content.trim() && msg.content !== '(无回复)') {
@@ -983,10 +1009,11 @@ async function sendMessage() {
             // Incremental streaming token
             currentText += evt.content;
             if (_isV2Mode) {
-              // In V2 mode, text_delta contains raw XML — store separately, not as user-facing content
+              // In V2 mode, text_delta contains raw XML — store separately
               _v2RawXml = currentText;
               // Show V2 reasoning text as streaming content (if available)
-              state.messages[msgIdx]._streamingText = _v2ReasoningText || '';
+              // If no reasoning yet, show a stripped version of raw text so user sees real-time progress
+              state.messages[msgIdx]._streamingText = _v2ReasoningText || _stripXmlTags(currentText);
             } else {
               // V1 mode: text_delta IS the user-facing content
               state.messages[msgIdx]._streamingText = currentText;
@@ -1083,6 +1110,8 @@ async function sendMessage() {
             _isV2Mode = true;
           } else if (evt.type === 'v2_output_parsed') {
             // LLM output was parsed into structured format
+            // IMPORTANT: Also mark V2 mode here — some server configs may skip v2_context
+            _isV2Mode = true;
             // evt.data contains: {usersays_correct, task_plan, tools_to_call, remember, recall, ask_user, finish}
             // Store for rendering
             state.messages[msgIdx]._v2Parsed = evt.data;
@@ -1163,9 +1192,10 @@ async function sendMessage() {
             if (!state.messages[msgIdx]._v2Reasoning) {
               state.messages[msgIdx]._v2Reasoning = '';
             }
-            // Add separator between multiple reasoning events (e.g., beforecalltext + final text)
-            if (state.messages[msgIdx]._v2Reasoning.length > 0 && evt.content) {
-              state.messages[msgIdx]._v2Reasoning += '\n\n';
+            // Smart separator: only add newline if previous reasoning doesn't already end with one
+            var prev = state.messages[msgIdx]._v2Reasoning;
+            if (prev.length > 0 && evt.content && !prev.endsWith('\n') && !evt.content.startsWith('\n')) {
+              state.messages[msgIdx]._v2Reasoning += '\n';
             }
             state.messages[msgIdx]._v2Reasoning += evt.content;
             _v2ReasoningText = state.messages[msgIdx]._v2Reasoning;