diff --git a/agent/deep_assistant.py b/agent/deep_assistant.py
index 49640e5..7933792 100644
--- a/agent/deep_assistant.py
+++ b/agent/deep_assistant.py
@@ -166,7 +166,38 @@ async def init_agent(config: AgentConfig):
         "api_key": config.api_key
     }
     if config.generate_cfg:
-        model_kwargs.update(config.generate_cfg)
+        # 内部使用的参数，不应传给任何 LLM
+        internal_params = {
+            'tool_output_max_length',
+            'tool_output_truncation_strategy',
+            'tool_output_filters',
+            'tool_output_exclude',
+            'preserve_code_blocks',
+            'preserve_json',
+        }
+
+        # Anthropic 不支持的 OpenAI 特有参数
+        openai_only_params = {
+            'n',  # 生成多少个响应
+            'presence_penalty',
+            'frequency_penalty',
+            'logprobs',
+            'top_logprobs',
+            'logit_bias',
+            'seed',
+            'suffix',
+            'best_of',
+            'echo',
+            'user',
+        }
+
+        # 根据提供商决定需要过滤的参数
+        params_to_filter = internal_params.copy()
+        if model_provider == 'anthropic':
+            params_to_filter.update(openai_only_params)
+
+        filtered_cfg = {k: v for k, v in config.generate_cfg.items() if k not in params_to_filter}
+        model_kwargs.update(filtered_cfg)
     llm_instance = init_chat_model(**model_kwargs)
 
     # 创建新的 agent（不再缓存）
diff --git a/agent/mem0_manager.py b/agent/mem0_manager.py
index 1b73833..b108d81 100644
--- a/agent/mem0_manager.py
+++ b/agent/mem0_manager.py
@@ -218,6 +218,38 @@ class Mem0Manager:
         """
         return self._sync_pool
 
+    def _cleanup_mem0_instance(self, mem0_instance: Any) -> None:
+        """清理 Mem0 实例，释放数据库连接
+
+        Mem0 的 PGVector 实现在初始化时获取连接并持有，
+        只有在 __del__ 时才归还。Python 的 GC 不保证 __del__ 立即被调用，
+        可能导致连接池耗尽。此方法显式释放连接。
+
+        Args:
+            mem0_instance: Mem0 Memory 实例
+        """
+        try:
+            # Mem0 Memory 实例有一个 vector_store 属性，类型是 PGVector
+            if hasattr(mem0_instance, 'vector_store'):
+                vector_store = mem0_instance.vector_store
+                # PGVector 有 conn 和 connection_pool 属性
+                if hasattr(vector_store, 'conn') and hasattr(vector_store, 'connection_pool'):
+                    if vector_store.connection_pool is not None:
+                        try:
+                            # 先关闭游标
+                            if hasattr(vector_store, 'cur') and vector_store.cur:
+                                vector_store.cur.close()
+                            # 归还连接到池
+                            vector_store.connection_pool.putconn(vector_store.conn)
+                            # 标记为已清理，防止 __del__ 重复释放
+                            vector_store.conn = None
+                            vector_store.connection_pool = None
+                            logger.debug("Successfully released Mem0 database connection back to pool")
+                        except Exception as e:
+                            logger.warning(f"Error releasing Mem0 connection: {e}")
+        except Exception as e:
+            logger.warning(f"Error cleaning up Mem0 instance: {e}")
+
     async def get_mem0(
         self,
         user_id: str,
@@ -249,8 +281,10 @@ class Mem0Manager:
 
         # 检查缓存大小，超过则移除最旧的
         if len(self._instances) >= self._max_instances:
-            removed_key, _ = self._instances.popitem(last=False)
-            logger.debug(f"Mem0 instance cache full, removed oldest entry: {removed_key}")
+            removed_key, removed_instance = self._instances.popitem(last=False)
+            # 显式释放连接，避免等待 GC 导致连接池耗尽
+            self._cleanup_mem0_instance(removed_instance)
+            logger.debug(f"Mem0 instance cache full, removed and cleaned: {removed_key}")
 
         # 创建新实例
         mem0_instance = await self._create_mem0_instance(