适配记忆注入与全局记忆

.astrbot-plugin/i18n/en-US.json

@@ -0,0 +1,69 @@
+{
+  "metadata": {
+    "display_name": "Simple Long Memory",
+    "short_desc": "Knowledge-base backed long-term memory, recall, and management tools.",
+    "desc": "Adds long-term memory to AstrBot using the built-in knowledge base for user preferences, past interactions, and important facts."
+  },
+  "config": {
+    "kb_name": {
+      "description": "Memory Knowledge Base",
+      "hint": "Select the knowledge base used to store long-term memories. Required. Create one first in Knowledge Base Management and configure an embedding model."
+    },
+    "extraction_provider_id": {
+      "description": "Memory Extraction Model",
+      "hint": "LLM used to extract memories from conversations. Leave empty to use the current chat model."
+    },
+    "summarization_provider_id": {
+      "description": "Memory Summarization Model",
+      "hint": "LLM used for future memory merging and compression. Leave empty to use the current chat model."
+    },
+    "auto_memorize": {
+      "description": "Automatic Memory Mode",
+      "hint": "When enabled, memories are automatically extracted and stored after conversations."
+    },
+    "extraction_interval": {
+      "description": "Extraction Interval (turns)",
+      "hint": "Run memory extraction every N conversation turns and summarize those turns."
+    },
+    "extraction_min_content_length": {
+      "description": "Minimum Extraction Length",
+      "hint": "Skip memory extraction when the accumulated conversation is shorter than this many characters."
+    },
+    "global_memory": {
+      "description": "Cross-Session Personal Recall",
+      "hint": "When enabled, personal memories can be recalled across sessions for the same user. When disabled, only personal memories from the current session are recalled."
+    },
+    "max_memories_per_inject": {
+      "description": "Memories Per Injection",
+      "hint": "Maximum number of memories injected into each LLM request."
+    },
+    "max_memory_list_scan": {
+      "description": "Memory List Scan Limit",
+      "hint": "Limits how many records /memory list scans to calculate memories visible to the current user. The actual scan amount is derived from the page and page size and will not exceed this limit."
+    },
+    "memory_delete_scan_page_size": {
+      "description": "Delete Scan Page Size",
+      "hint": "Number of matched records scanned per page before deleting or clearing memories. Larger values are faster but use more memory per operation."
+    },
+    "memory_ttl_days": {
+      "description": "Memory TTL (days)",
+      "hint": "Low-frequency memories older than this many days may be compressed into impressions."
+    },
+    "use_reranker": {
+      "description": "Enable Reranker",
+      "hint": "Use the knowledge base reranker to improve memory recall ordering. Requires a reranker model configured on the knowledge base."
+    },
+    "optimize_recall_query": {
+      "description": "Optimize Recall Query",
+      "hint": "Before each memory recall, call the extraction model to extract search keywords for better recall accuracy. This runs on every conversation, so use a fast lightweight model."
+    },
+    "optimize_recall_query_timeout": {
+      "description": "Recall Optimization Timeout (seconds)",
+      "hint": "Maximum wait time for recall query optimization. On timeout, the original query is used so conversation responses are not blocked."
+    },
+    "enable_admin_global_memory_tool": {
+      "description": "Enable Admin Global Memory Tool",
+      "hint": "When enabled, provides the memory_store_global tool so administrators can ask AI to add global memories that apply to all chats."
+    }
+  }
+}

.astrbot-plugin/i18n/zh-CN.json

@@ -0,0 +1,69 @@
+{
+  "metadata": {
+    "display_name": "简单长期记忆",
+    "short_desc": "基于知识库的长期记忆、召回和管理工具。",
+    "desc": "为 AstrBot 提供长期记忆能力，基于内置知识库实现用户偏好、历史交互和重要事实的记忆存储与召回。"
+  },
+  "config": {
+    "kb_name": {
+      "description": "记忆知识库",
+      "hint": "选择用于存储长期记忆的知识库（必填，请先在知识库管理中创建，知识库需配置嵌入模型）"
+    },
+    "extraction_provider_id": {
+      "description": "记忆提取模型",
+      "hint": "用于从对话中提取记忆的 LLM 模型（留空则使用会话主 LLM）"
+    },
+    "summarization_provider_id": {
+      "description": "记忆总结模型",
+      "hint": "用于记忆合并、压缩等总结操作的 LLM 模型（留空则使用会话主 LLM）"
+    },
+    "auto_memorize": {
+      "description": "自动记忆模式",
+      "hint": "开启后会在对话后自动提取并存储记忆"
+    },
+    "extraction_interval": {
+      "description": "记忆提取间隔（轮）",
+      "hint": "每 N 轮对话后触发一次记忆提取，对这 N 轮对话内容进行总结"
+    },
+    "extraction_min_content_length": {
+      "description": "最小提取内容长度",
+      "hint": "对话总内容低于此字符数时跳过记忆提取，避免无意义短对话被记忆"
+    },
+    "global_memory": {
+      "description": "跨会话召回个人记忆",
+      "hint": "开启后同一用户可跨会话召回个人记忆；关闭后仅召回当前会话的个人记忆"
+    },
+    "max_memories_per_inject": {
+      "description": "每次注入的记忆数量",
+      "hint": "LLM 请求时注入的最大记忆条数"
+    },
+    "max_memory_list_scan": {
+      "description": "记忆列表扫描上限",
+      "hint": "限制 /memory list 为计算当前用户可见记忆而扫描的最大记录数。实际扫描量会按当前页码和每页数量推导，并不会超过此上限"
+    },
+    "memory_delete_scan_page_size": {
+      "description": "记忆删除扫描分页大小",
+      "hint": "删除或清空记忆前，每页扫描多少条匹配记录用于同步清理知识库文档记录。数值越大速度越快但单次内存占用更高"
+    },
+    "memory_ttl_days": {
+      "description": "记忆生命周期（天）",
+      "hint": "超过此天数的低频记忆将被压缩为印象"
+    },
+    "use_reranker": {
+      "description": "启用重排序",
+      "hint": "记忆召回时使用知识库配置的重排序模型优化结果排序。需先在知识库设置中配置重排序模型，未配置时此选项无效"
+    },
+    "optimize_recall_query": {
+      "description": "启用检索优化",
+      "hint": "开启后在每次记忆召回前调用记忆提取模型提炼检索关键词，提高召回准确率。每次对话都会调用，建议配置响应速度快的轻量模型"
+    },
+    "optimize_recall_query_timeout": {
+      "description": "检索优化超时（秒）",
+      "hint": "限制检索优化模型调用的最长等待时间。超时后将跳过优化并使用原始检索内容，避免阻塞对话响应"
+    },
+    "enable_admin_global_memory_tool": {
+      "description": "启用管理员全局记忆工具",
+      "hint": "开启后提供 memory_store_global 工具，仅管理员可通过 AI 添加对所有会话生效的全局记忆"
+    }
+  }
+}

CHANGELOG.md

@@ -1,5 +1,15 @@
 # Changelog
 
+## v0.3.2 (2026-05-14)
+
+### 新增
+- 新增 AstrBot 插件 i18n 配置，提供 `zh-CN` / `en-US` 双语 WebUI 元数据和配置项文案。
+- 新增 `enable_admin_global_memory_tool` 配置和 `memory_store_global` LLM 工具。开启后管理员可指挥 AI 写入 `global` 作用域记忆，后续所有会话都会参与召回。
+
+### 变更
+- 记忆注入优先使用 AstrBot v4.24+ 的临时用户内容区，并标记为本轮临时内容，避免写入会话历史；旧版回退到最早的 `user` 上下文位置，不再拼接到当前 prompt 前。
+- 记忆注入包裹说明强化为“长期记忆检索参考”，明确不是当前正在发生的事情，也不是用户指令；同时在顶部提醒 AI 必要时使用 `memory_recall(query)` 工具继续搜索更多记忆。
+
 ## v0.3.1 (2026-05-03)
 
 ### 新增

README.md

@@ -41,6 +41,7 @@ https://github.com/piexian/astrbot_plugin_simple_long_memory
 | use_reranker | 记忆召回时启用重排序（需知识库已配置重排序模型） | `true` |
 | optimize_recall_query | 启用检索优化（LLM 提炼关键词） | `false` |
 | optimize_recall_query_timeout | 检索优化超时（秒） | `10` |
+| enable_admin_global_memory_tool | 启用管理员全局记忆工具 | `false` |
 
 ## 使用方法
 
@@ -68,6 +69,7 @@ https://github.com/piexian/astrbot_plugin_simple_long_memory
 
 | 作用域 | 说明 | 日常例子 |
 |--------|------|----------|
+| `global` | 全局记忆，所有会话可召回（仅管理员工具可写入） | "机器人回答某项目问题时优先使用内部术语表" |
 | `personal` | 个人记忆，仅自己可见 | "我比较喜欢喝拿铁"、"下周要出差" |
 | `group` | 群共享记忆，群友都可见 | "群里约了每周五打游戏"、"这个群的固定梗" |
 | `conversation` | 当前会话临时上下文 | "刚才说的那个 bug 还没修完" |
@@ -109,6 +111,7 @@ AI 可以通过以下工具主动操作记忆：
 
 - `memory_recall(query)` — 搜索长期记忆
 - `memory_store(content, memory_type, disclosure)` — 存储记忆
+- `memory_store_global(content, memory_type, disclosure)` — 存储全局记忆（需开启 `enable_admin_global_memory_tool`，仅管理员可用）
 - `memory_forget(uri)` — 删除记忆
 
 ### 记忆类型
@@ -122,7 +125,7 @@ AI 可以通过以下工具主动操作记忆：
 
 ## 工作原理
 
-1. **记忆注入**：在每次 LLM 请求前，根据用户输入通过 embedding 检索召回相关记忆，以 `user` 角色注入到对话上下文顶部（不占用 system prompt）
+1. **记忆注入**：在每次 LLM 请求前，根据用户输入通过 embedding 检索召回相关记忆；AstrBot v4.24+ 优先注入到临时用户内容区（仅本轮请求生效），旧版回退到最早的 `user` 上下文位置（不占用 system prompt，不覆盖当前输入）
 2. **自动提取**：每隔 `extraction_interval` 轮对话，将累积的对话内容发送给 LLM 提取值得记忆的信息并自动存储
 3. **用户隔离**：所有记忆操作通过 metadata 中的 `user_id` 字段过滤，确保用户间记忆完全隔离
 4. **记忆存储**：记忆以向量形式存储在知识库中，支持语义检索

_conf_schema.json

@@ -90,5 +90,11 @@
     "default": 10,
     "hint": "限制检索优化模型调用的最长等待时间。超时后将跳过优化并使用原始检索内容，避免阻塞对话响应",
     "slider": {"min": 1, "max": 20, "step": 1}
+  },
+  "enable_admin_global_memory_tool": {
+    "description": "启用管理员全局记忆工具",
+    "type": "bool",
+    "default": false,
+    "hint": "开启后提供 memory_store_global 工具，仅管理员可通过 AI 添加对所有会话生效的全局记忆"
   }
 }

main.py

@@ -65,7 +65,13 @@ def _sanitize_string_list(value: Any, limit: int = 8) -> list[str]:
 
 
 def _normalize_extracted_scope(scope: str, session_type: str) -> str:
+    """规范化自动提取的记忆作用域。
+
+    自动提取不能写入全局记忆；全局记忆只能由管理员工具显式创建。
+    """
     scope = normalize_memory_scope(scope)
+    if scope == MemoryScope.GLOBAL:
+        return MemoryScope.PERSONAL
     if session_type != "group" and scope == MemoryScope.GROUP:
         return MemoryScope.PERSONAL
     return scope
@@ -133,6 +139,54 @@ def _build_recall_query(prompt: str, contexts: list[dict[str, Any]]) -> str:
     return "\n".join(parts)
 
 
+def _make_memory_text_part(content: str) -> Any | None:
+    """构造 AstrBot 用户内容块；新版支持标记为本轮临时内容。"""
+    try:
+        from astrbot.core.agent.message import TextPart
+    except Exception:
+        return None
+
+    part = TextPart(text=content)
+    mark_as_temp = getattr(part, "mark_as_temp", None)
+    if callable(mark_as_temp):
+        marked_part = mark_as_temp()
+        if marked_part is not None:
+            part = marked_part
+    return part
+
+
+def _append_extra_user_content(request: ProviderRequest, content: str) -> bool:
+    """优先使用 AstrBot 动态用户内容块注入记忆。"""
+    if not hasattr(request, "extra_user_content_parts"):
+        return False
+
+    parts = getattr(request, "extra_user_content_parts", None)
+    if parts is None:
+        parts = []
+        request.extra_user_content_parts = parts
+    if not hasattr(parts, "append"):
+        return False
+
+    part = _make_memory_text_part(content)
+    if part is None:
+        return False
+
+    parts.append(part)
+    return True
+
+
+def _inject_memory_context(request: ProviderRequest, content: str) -> str:
+    """注入记忆上下文，返回实际使用的注入位置。"""
+    if _append_extra_user_content(request, content):
+        return "extra_user_content_parts"
+
+    contexts = _normalize_contexts(getattr(request, "contexts", None))
+    memory_msg = {"role": "user", "content": content}
+    # 旧版回退：放在 contexts 最前面，使当前 prompt 仍保持最后、优先级更高。
+    request.contexts = [memory_msg] + contexts
+    return "contexts 顶部"
+
+
 def _clamp_timeout(
     value: Any, default: int = DEFAULT_RECALL_QUERY_OPTIMIZATION_TIMEOUT
 ) -> int:
@@ -770,22 +824,10 @@ async def inject_memories(self, event: AstrMessageEvent, request: ProviderReques
                 # format_memory_for_injection 已返回含 <user_context_reference> 包装的完整字符串
                 safe_memory_context = format_memory_for_injection(memories)
                 if safe_memory_context:
-                    # 优先注入到 contexts 顶部（如果存在）
-                    # 使用 user 角色而非 system，降低优先级
-                    if contexts:
-                        memory_msg = {"role": "user", "content": safe_memory_context}
-                        request.contexts = [memory_msg] + contexts
-                        logger.debug(
-                            f"[简单长期记忆] 注入 {len(memories)} 条记忆到 contexts 顶部"
-                        )
-                    else:
-                        # 回退：注入到 prompt 前面
-                        request.prompt = (
-                            f"{safe_memory_context}\n\n{request.prompt or ''}"
-                        )
-                        logger.debug(
-                            f"[简单长期记忆] 注入 {len(memories)} 条记忆到 prompt 前"
-                        )
+                    inject_target = _inject_memory_context(request, safe_memory_context)
+                    logger.debug(
+                        f"[简单长期记忆] 注入 {len(memories)} 条记忆到 {inject_target}"
+                    )
 
         except Exception as e:
             logger.warning(f"[简单长期记忆] 注入记忆失败: {e}")
@@ -1459,6 +1501,54 @@ async def tool_store(
         )
         return f"Memory stored: {uri}"
 
+    @filter.llm_tool(name="memory_store_global")
+    async def tool_store_global(
+        self,
+        event: AstrMessageEvent,
+        content: str,
+        memory_type: str = "fact",
+        disclosure: str = "",
+    ) -> str:
+        """Store a global long-term memory visible to all chats.
+
+        Only use this when the current user is an admin and explicitly asks to
+        add a global memory, global rule, shared fact, or bot-wide preference.
+
+        Args:
+            content(string): global content to remember for all users/chats
+            memory_type(string): memory type (fact/preference/event/context)
+            disclosure(string): condition description for triggering recall
+        """
+        if not self.config.get("enable_admin_global_memory_tool", False):
+            return "Global memory tool is disabled in plugin config"
+
+        if not event.is_admin():
+            return "Only administrators can store global memories"
+
+        if not self.memory_mgr:
+            return "Memory plugin not initialized"
+
+        content = _sanitize_memory_content(content)
+        if not content:
+            return "Invalid memory content"
+
+        domain = normalize_domain(memory_type)
+        uri = str(MemoryURI.generate(domain))
+
+        await self.memory_mgr.store_memory(
+            event=event,
+            content=content,
+            domain=domain,
+            uri=uri,
+            memory_type=MemoryType.PERMANENT,
+            disclosure=_sanitize_memory_content(disclosure)[:200] if disclosure else "",
+            importance=5,
+            memory_scope=MemoryScope.GLOBAL,
+            visibility="group",
+            subject="global",
+        )
+        return f"Global memory stored: {uri}"
+
     @filter.llm_tool(name="memory_forget")
     async def tool_forget(self, event: AstrMessageEvent, uri: str) -> str:
         """Delete a specific memory by URI