diff --git a/astrbot/core/config/default.py b/astrbot/core/config/default.py
index 4a6364979..648b2bc9a 100644
--- a/astrbot/core/config/default.py
+++ b/astrbot/core/config/default.py
@@ -50,6 +50,7 @@ DEFAULT_CONFIG = {
         "default_personality": "default",
         "prompt_prefix": "",
         "max_context_length": -1,
+        "dequeue_context_length": 1,
         "streaming_response": False,
     },
     "provider_stt_settings": {
@@ -997,6 +998,11 @@ CONFIG_METADATA_2 = {
                         "type": "int",
                         "hint": "超出这个数量时将丢弃最旧的部分，用户和AI的一轮聊天记为 1 条。-1 表示不限制，默认为不限制。",
                     },
+                    "dequeue_context_length": {
+                        "description": "丢弃对话数量(条)",
+                        "type": "int",
+                        "hint": "超出 最多携带对话数量(条) 时，丢弃多少条记录，用户和AI的一轮聊天记为 1 条。适宜的配置，可以提高超长上下文对话 deepseek 命中缓存效果，理想情况下计费将降低到1/3以下",
+                    },
                     "streaming_response": {
                         "description": "启用流式回复",
                         "type": "bool",
diff --git a/astrbot/core/pipeline/process_stage/method/llm_request.py b/astrbot/core/pipeline/process_stage/method/llm_request.py
index a9f793626..547099e21 100644
--- a/astrbot/core/pipeline/process_stage/method/llm_request.py
+++ b/astrbot/core/pipeline/process_stage/method/llm_request.py
@@ -38,6 +38,10 @@ class LLMRequestSubStage(Stage):
         self.max_context_length = ctx.astrbot_config["provider_settings"][
             "max_context_length"
         ]  # int
+        self.dequeue_context_length = min(
+            max(1,ctx.astrbot_config["provider_settings"]["dequeue_context_length"]),
+            self.max_context_length - 1
+        )  # int
         self.streaming_response = ctx.astrbot_config["provider_settings"][
             "streaming_response"
         ]  # bool
@@ -137,7 +141,7 @@ class LLMRequestSubStage(Stage):
             and len(req.contexts) // 2 > self.max_context_length
         ):
             logger.debug("上下文长度超过限制，将截断。")
-            req.contexts = req.contexts[-self.max_context_length * 2 :]
+            req.contexts = req.contexts[-(self.max_context_length - self.dequeue_context_length) * 2 :]
 
         # session_id
         if not req.session_id: