diff --git a/plugins/ai_auto_response/context/context_builder.py b/plugins/ai_auto_response/context/context_builder.py
index 230f2c4..9316d31 100644
--- a/plugins/ai_auto_response/context/context_builder.py
+++ b/plugins/ai_auto_response/context/context_builder.py
@@ -53,6 +53,8 @@ class ContextBuilder:
             "speaker_name_clean": self._clean_display_name(sender_name),
             "is_at": bool(trigger.get("is_at", False)),
             "is_directed": bool(trigger.get("is_directed", False)),
+            # 这类标记会被后面的 prompt 策略层消费，用来决定要不要放开群级记忆。
+            "is_group_memory_query": bool(trigger.get("is_group_memory_query", False)),
             "recent_message_items": self._build_recent_message_items(selected_messages),
             "recent_messages": recent_lines,
             "recent_summary": "",
diff --git a/plugins/ai_auto_response/core/response_planner.py b/plugins/ai_auto_response/core/response_planner.py
index cfe44dd..46ca1b4 100644
--- a/plugins/ai_auto_response/core/response_planner.py
+++ b/plugins/ai_auto_response/core/response_planner.py
@@ -9,6 +9,12 @@ class ResponsePlanner:
         # 而是一句短短的回怼或挡回去，所以这里强制走 social_short。
         if trigger.get("is_directed_abuse"):
             return "social_short"
+        # “群里最近都聊什么”这类问题，本质是在问群记忆摘要：
+        # 1. 如果继续走 qa_fast，就只会优先依赖最近现场消息；
+        # 2. 这里直接抬到 qa_with_context，后面才会打开群事实/向量记忆等补充层；
+        # 3. 不依赖 flow_state，是因为这类问题和当前场子热不热关系不大。
+        if trigger.get("is_question") and trigger.get("is_group_memory_query"):
+            return "qa_with_context"
         if trigger.get("is_question"):
             return "qa_with_context" if flow_state in {"engaged", "deep_engaged"} else "qa_fast"
         if trigger.get("is_followup"):
diff --git a/plugins/ai_auto_response/core/triggers.py b/plugins/ai_auto_response/core/triggers.py
index 896bff0..6845a64 100644
--- a/plugins/ai_auto_response/core/triggers.py
+++ b/plugins/ai_auto_response/core/triggers.py
@@ -11,6 +11,16 @@ QUESTION_PATTERNS = [
     r"\?$", r"？$", r"怎么", r"如何", r"咋弄", r"为啥", r"为什么",
     r"有人知道", r"谁知道", r"能不能", r"可以吗", r"报错", r"怎么解决",
 ]
+GROUP_MEMORY_QUERY_PATTERNS = [
+    # 这组模式专门识别“回顾群里最近在聊什么”的提问：
+    # 1. 这类问题本质上不是要 bot 现场接一句，而是要它调动群级记忆来做概括；
+    # 2. 之前它会被当成普通问句，落进 qa_fast，结果只看最近 30 条现场消息；
+    # 3. 单独打标后，后续策略层才能有依据切到 qa_with_context。
+    r"(最近|这两天|这几天|近期).*(聊|讨论|在说|在聊|话题|主题|重点|近况)",
+    r"(都|最近).*(聊什么|说什么|讨论什么|在聊啥|在说啥|啥话题)",
+    r"(群里|大家|他们).*(最近|这两天|这几天).*(聊|讨论|话题|重点)",
+    r"(总结|概括).*(最近|这两天|这几天|近期).*(聊天|讨论|话题|内容)",
+]
 SOCIAL_PATTERNS = [r"小牛", r"在吗", r"出来", r"帮忙看", r"看看"]
 LIGHT_SOCIAL_PATTERNS = [
     r"哈哈", r"笑死", r"绷不住", r"离谱", r"逆天", r"牛逼", r"卧槽", r"我去",
@@ -34,6 +44,9 @@ class TriggerResult:
     is_directed_abuse: bool = False
     is_followup: bool = False
     is_social_call: bool = False
+    # 这类问题是在问“群里最近都在聊啥”，
+    # 需要优先动用群事实/长期摘要/向量召回，而不是只看短期现场窗口。
+    is_group_memory_query: bool = False
     is_returning_member: bool = False
     should_respond: bool = False
     topic: str = ""
@@ -94,6 +107,11 @@ class TriggerRouter:
                     result.is_directed = True
                     result.reasons.append("question_named_bot")
                 result.reasons.append("question")
+        if self._is_group_memory_query(content):
+            # “最近都聊什么”这类提问单独记下来，供后面的回复模式和记忆策略升级。
+            # 这里不直接改 should_respond，是为了继续遵守“必须是定向提问才回答”的总原则。
+            result.is_group_memory_query = True
+            result.reasons.append("group_memory_query")
         if is_directed_abuse(
             content,
             directed=bool(message.get("is_at")) or named_to_bot or conversation_hints.get("quote_targets_bot", False),
@@ -186,6 +204,12 @@ class TriggerRouter:
                 return keyword
         return ""
 
+    def _is_group_memory_query(self, content: str) -> bool:
+        text = str(content or "").strip()
+        if not text:
+            return False
+        return any(re.search(pattern, text, flags=re.IGNORECASE) for pattern in GROUP_MEMORY_QUERY_PATTERNS)
+
     def _has_bot_name(self, content_lower: str) -> bool:
         text = str(content_lower or "").strip().lower()
         if not text:
diff --git a/plugins/ai_auto_response/main.py b/plugins/ai_auto_response/main.py
index f710c1c..73142e5 100644
--- a/plugins/ai_auto_response/main.py
+++ b/plugins/ai_auto_response/main.py
@@ -1196,6 +1196,7 @@ class AIAutoResponsePlugin(MessagePluginInterface):
         trigger_type = str(context.get("trigger_type", "none") or "none")
         is_at = bool(context.get("is_at", False))
         is_directed = bool(context.get("is_directed", False))
+        is_group_memory_query = bool(context.get("is_group_memory_query", False))
         is_followup = bool(memory_hints.get("is_followup", False))
         returning_state = str(memory_hints.get("returning_member_state", "") or "").strip()
         strong_directed = is_at or is_directed or trigger_type in {"at_trigger", "quote_followup_trigger"}
@@ -1231,9 +1232,15 @@ class AIAutoResponsePlugin(MessagePluginInterface):
         # 1. 用户希望回答能带上群里的长期背景和互动关系；
         # 2. 关系记忆仍会经过相关性过滤，所以放宽入口不会直接把无关关系灌进去；
         # 3. 这样技术问答里也更容易利用“谁经常和谁接话、谁常问哪类问题”的弱背景。
-        allow_social_memory = is_question_like
-        allow_group_facts = reply_mode == "qa_with_context"
-        allow_vector_memory = reply_mode == "qa_with_context" or returning_state == "long_absent_member"
+        allow_social_memory = is_question_like or is_group_memory_query
+        # “最近都聊什么”这类问题，本身就是在问群级记忆，
+        # 所以哪怕当前只是普通问答入口，也要把群事实和向量层放开。
+        allow_group_facts = reply_mode == "qa_with_context" or is_group_memory_query
+        allow_vector_memory = (
+            reply_mode == "qa_with_context"
+            or returning_state == "long_absent_member"
+            or is_group_memory_query
+        )
 
         return {
             "target_reply_chars": target_reply_chars_map.get(reply_mode, 10),
@@ -1292,6 +1299,13 @@ class AIAutoResponsePlugin(MessagePluginInterface):
 
     @staticmethod
     def _is_text_relevant(content: str, memory_text: str) -> bool:
+        # 对“最近都聊什么”这类群聊回顾型问题做一个显式兜底：
+        # 1. 这类问题天然缺少技术关键词，严格按词重叠时经常会得到 0 命中；
+        # 2. 但它问的恰恰就是“群级记忆摘要”，不应该再被相关性门槛二次挡掉；
+        # 3. 因此只要当前问题像群聊话题回顾，而记忆文本也明显是群摘要/群事实，就直接放行。
+        if AIAutoResponsePlugin._looks_like_group_memory_query(content):
+            if AIAutoResponsePlugin._looks_like_group_memory_text(memory_text):
+                return True
         content_tokens = AIAutoResponsePlugin._extract_relevance_tokens(content)
         memory_tokens = AIAutoResponsePlugin._extract_relevance_tokens(memory_text)
         if not content_tokens or not memory_tokens:
@@ -1299,6 +1313,30 @@ class AIAutoResponsePlugin(MessagePluginInterface):
         overlap = content_tokens & memory_tokens
         return len(overlap) >= 1
 
+    @staticmethod
+    def _looks_like_group_memory_query(content: str) -> bool:
+        text = str(content or "").strip()
+        if not text:
+            return False
+        patterns = [
+            r"(最近|这两天|这几天|近期).*(聊|讨论|在说|在聊|话题|主题|重点|近况)",
+            r"(都|最近).*(聊什么|说什么|讨论什么|在聊啥|在说啥|啥话题)",
+            r"(群里|大家|他们).*(最近|这两天|这几天).*(聊|讨论|话题|重点)",
+            r"(总结|概括).*(最近|这两天|这几天|近期).*(聊天|讨论|话题|内容)",
+        ]
+        return any(re.search(pattern, text, flags=re.IGNORECASE) for pattern in patterns)
+
+    @staticmethod
+    def _looks_like_group_memory_text(memory_text: str) -> bool:
+        text = str(memory_text or "").strip()
+        if not text:
+            return False
+        markers = [
+            "群长期背景", "长期摘要", "稳定主题", "近期重点", "未决问题",
+            "群事实", "最近沉淀", "最近长期反复出现", "下面是群", "相关话题",
+        ]
+        return any(marker in text for marker in markers)
+
     @staticmethod
     def _extract_relevance_tokens(text: str) -> set[str]:
         raw = str(text or "").lower()