diff --git a/core/src/main/java/com/taobao/arthas/core/command/express/ExpressFactory.java b/core/src/main/java/com/taobao/arthas/core/command/express/ExpressFactory.java
index d2d578bc17f..886c910f1bb 100644
--- a/core/src/main/java/com/taobao/arthas/core/command/express/ExpressFactory.java
+++ b/core/src/main/java/com/taobao/arthas/core/command/express/ExpressFactory.java
@@ -1,5 +1,7 @@
 package com.taobao.arthas.core.command.express;
 
+import java.lang.ref.WeakReference;
+
 /**
  * ExpressFactory
  * @author ralf0131 2017-01-04 14:40.
@@ -7,12 +9,14 @@
  */
 public class ExpressFactory {
 
-    private static final ThreadLocal<Express> expressRef = new ThreadLocal<Express>() {
-        @Override
-        protected Express initialValue() {
-            return new OgnlExpress();
-        }
-    };
+    /**
+     * 这里不能直接在 ThreadLocalMap 里强引用 Express（它由 ArthasClassLoader 加载），否则 stop/detach 后会被业务线程持有，
+     * 导致 ArthasClassLoader 无法被 GC 回收。
+     *
+     * 用 WeakReference 打断强引用链：Thread -> ThreadLocalMap -> value(WeakReference) -X-> Express。
+     */
+    private static final ThreadLocal<WeakReference<Express>> expressRef = ThreadLocal
+            .withInitial(() -> new WeakReference<Express>(new OgnlExpress()));
 
     /**
      * get ThreadLocal Express Object
@@ -20,7 +24,13 @@ protected Express initialValue() {
      * @return
      */
     public static Express threadLocalExpress(Object object) {
-        return expressRef.get().reset().bind(object);
+        WeakReference<Express> reference = expressRef.get();
+        Express express = reference == null ? null : reference.get();
+        if (express == null) {
+            express = new OgnlExpress();
+            expressRef.set(new WeakReference<Express>(express));
+        }
+        return express.reset().bind(object);
     }
 
     public static Express unpooledExpress(ClassLoader classloader) {
@@ -29,4 +39,4 @@ public static Express unpooledExpress(ClassLoader classloader) {
         }
         return new OgnlExpress(new ClassLoaderClassResolver(classloader));
     }
-}
\ No newline at end of file
+}
diff --git a/core/src/main/java/com/taobao/arthas/core/command/monitor200/TimeTunnelAdviceListener.java b/core/src/main/java/com/taobao/arthas/core/command/monitor200/TimeTunnelAdviceListener.java
index 6c9faf7084f..7a088f16f04 100644
--- a/core/src/main/java/com/taobao/arthas/core/command/monitor200/TimeTunnelAdviceListener.java
+++ b/core/src/main/java/com/taobao/arthas/core/command/monitor200/TimeTunnelAdviceListener.java
@@ -21,12 +21,22 @@
  */
 public class TimeTunnelAdviceListener extends AdviceListenerAdapter {
     private static final Logger logger = LoggerFactory.getLogger(TimeTunnelAdviceListener.class);
-    private final ThreadLocal<ObjectStack> argsRef = new ThreadLocal<ObjectStack>() {
-        @Override
-        protected ObjectStack initialValue() {
-            return new ObjectStack(512);
-        }
-    };
+    /**
+     * 用 JDK 的 Object[] 做一个固定大小的 ring stack（只存业务对象），避免把 ArthasClassLoader 加载的 ObjectStack 放进
+     * 业务线程的 ThreadLocalMap 里，导致 stop/detach 后 ArthasClassLoader 无法被 GC 回收。
+     *
+     * <pre>
+     * 约定：
+     * - store[0] 存储 int[1] 的 pos（0..cap）
+     * - store[1..cap] 存储 args（Object[]）
+     * </pre>
+     */
+    private static final int ARGS_STACK_SIZE = 512;
+    private final ThreadLocal<Object[]> argsRef = ThreadLocal.withInitial(() -> {
+        Object[] store = new Object[ARGS_STACK_SIZE + 1];
+        store[0] = new int[1];
+        return store;
+    });
 
     private TimeTunnelCommand command;
     private CommandProcess process;
@@ -46,7 +56,7 @@ public TimeTunnelAdviceListener(TimeTunnelCommand command, CommandProcess proces
     @Override
     public void before(ClassLoader loader, Class<?> clazz, ArthasMethod method, Object target, Object[] args)
             throws Throwable {
-        argsRef.get().push(args);
+        pushArgs(args);
         threadLocalWatch.start();
     }
 
@@ -54,7 +64,7 @@ public void before(ClassLoader loader, Class<?> clazz, ArthasMethod method, Obje
     public void afterReturning(ClassLoader loader, Class<?> clazz, ArthasMethod method, Object target, Object[] args,
                                Object returnObject) throws Throwable {
         //取出入参时的 args，因为在函数执行过程中 args可能被修改
-        args = (Object[]) argsRef.get().pop();
+        args = popArgs();
         afterFinishing(Advice.newForAfterReturning(loader, clazz, method, target, args, returnObject));
     }
 
@@ -62,10 +72,46 @@ public void afterReturning(ClassLoader loader, Class<?> clazz, ArthasMethod meth
     public void afterThrowing(ClassLoader loader, Class<?> clazz, ArthasMethod method, Object target, Object[] args,
                               Throwable throwable) {
         //取出入参时的 args，因为在函数执行过程中 args可能被修改
-        args = (Object[]) argsRef.get().pop();
+        args = popArgs();
         afterFinishing(Advice.newForAfterThrowing(loader, clazz, method, target, args, throwable));
     }
 
+    private void pushArgs(Object[] args) {
+        Object[] store = argsRef.get();
+        int[] posHolder = (int[]) store[0];
+
+        int cap = store.length - 1;
+        int pos = posHolder[0];
+        if (pos < cap) {
+            pos++;
+        } else {
+            // if stack is full, reset pos
+            pos = 1;
+        }
+        store[pos] = args;
+        posHolder[0] = pos;
+    }
+
+    private Object[] popArgs() {
+        Object[] store = argsRef.get();
+        int[] posHolder = (int[]) store[0];
+
+        int cap = store.length - 1;
+        int pos = posHolder[0];
+        if (pos > 0) {
+            Object[] args = (Object[]) store[pos];
+            store[pos] = null;
+            posHolder[0] = pos - 1;
+            return args;
+        }
+
+        pos = cap;
+        Object[] args = (Object[]) store[pos];
+        store[pos] = null;
+        posHolder[0] = pos - 1;
+        return args;
+    }
+
     private void afterFinishing(Advice advice) {
         double cost = threadLocalWatch.costInMillis();
         TimeFragment timeTunnel = new TimeFragment(advice, LocalDateTime.now(), cost);
@@ -103,57 +149,4 @@ private void afterFinishing(Advice advice) {
             abortProcess(process, command.getNumberOfLimit());
         }
     }
-
-    /**
-     * 
-     * <pre>
-     * 一个特殊的stack，为了追求效率，避免扩容。
-     * 因为这个stack的push/pop 并不一定成对调用，比如可能push执行了，但是后面的流程被中断了，pop没有被执行。
-     * 如果不固定大小，一直增长的话，极端情况下可能应用有内存问题。
-     * 如果到达容量，pos会重置，循环存储数据。所以使用这个Stack如果在极端情况下统计的数据会不准确，只用于monitor/watch等命令的计时。
-     * 
-     * </pre>
-     * 
-     * @author hengyunabc 2020-05-20
-     *
-     */
-    static class ObjectStack {
-        private Object[] array;
-        private int pos = 0;
-        private int cap;
-
-        public ObjectStack(int maxSize) {
-            array = new Object[maxSize];
-            cap = array.length;
-        }
-
-        public int size() {
-            return pos;
-        }
-
-        public void push(Object value) {
-            if (pos < cap) {
-                array[pos++] = value;
-            } else {
-                // if array is full, reset pos
-                pos = 0;
-                array[pos++] = value;
-            }
-        }
-
-        public Object pop() {
-            if (pos > 0) {
-                pos--;
-                Object object = array[pos];
-                array[pos] = null;
-                return object;
-            } else {
-                pos = cap;
-                pos--;
-                Object object = array[pos];
-                array[pos] = null;
-                return object;
-            }
-        }
-    }
 }
diff --git a/core/src/main/java/com/taobao/arthas/core/shell/command/internal/TermHandler.java b/core/src/main/java/com/taobao/arthas/core/shell/command/internal/TermHandler.java
index a8b3dae2212..fc2d29662d0 100644
--- a/core/src/main/java/com/taobao/arthas/core/shell/command/internal/TermHandler.java
+++ b/core/src/main/java/com/taobao/arthas/core/shell/command/internal/TermHandler.java
@@ -19,4 +19,4 @@ public String apply(String data) {
         term.write(data);
         return data;
     }
-}
+}
\ No newline at end of file
diff --git a/core/src/main/java/com/taobao/arthas/core/util/ThreadLocalWatch.java b/core/src/main/java/com/taobao/arthas/core/util/ThreadLocalWatch.java
index 9a5e1e257c8..a47657a5e15 100644
--- a/core/src/main/java/com/taobao/arthas/core/util/ThreadLocalWatch.java
+++ b/core/src/main/java/com/taobao/arthas/core/util/ThreadLocalWatch.java
@@ -8,25 +8,31 @@
  */
 public class ThreadLocalWatch {
 
-    private final ThreadLocal<LongStack> timestampRef = new ThreadLocal<LongStack>() {
-        @Override
-        protected LongStack initialValue() {
-            return new LongStack(1024 * 4);
-        }
-    };
+    /**
+     * 用 long[] 做一个固定大小的 ring stack，避免把 ArthasClassLoader 加载的对象塞到业务线程的 ThreadLocalMap 里，
+     * 从而在 stop/detach 后导致 ArthasClassLoader 无法被 GC 回收。
+     *
+     * <pre>
+     * 约定：
+     * - stack[0] 存储当前 pos（0..cap）
+     * - stack[1..cap] 存储数据
+     * </pre>
+     */
+    private static final int DEFAULT_STACK_SIZE = 1024 * 4;
+    private final ThreadLocal<long[]> timestampRef = ThreadLocal.withInitial(() -> new long[DEFAULT_STACK_SIZE + 1]);
 
     public long start() {
         final long timestamp = System.nanoTime();
-        timestampRef.get().push(timestamp);
+        push(timestampRef.get(), timestamp);
         return timestamp;
     }
 
     public long cost() {
-        return (System.nanoTime() - timestampRef.get().pop());
+        return (System.nanoTime() - pop(timestampRef.get()));
     }
 
     public double costInMillis() {
-        return (System.nanoTime() - timestampRef.get().pop()) / 1000000.0;
+        return (System.nanoTime() - pop(timestampRef.get())) / 1000000.0;
     }
 
     /**
@@ -42,39 +48,31 @@ public double costInMillis() {
      * @author hengyunabc 2019-11-20
      *
      */
-    static class LongStack {
-        private long[] array;
-        private int pos = 0;
-        private int cap;
-
-        public LongStack(int maxSize) {
-            array = new long[maxSize];
-            cap = array.length;
-        }
-
-        public int size() {
-            return pos;
+    static void push(long[] stack, long value) {
+        int cap = stack.length - 1;
+        int pos = (int) stack[0];
+        if (pos < cap) {
+            pos++;
+        } else {
+            // if stack is full, reset pos
+            pos = 1;
         }
+        stack[pos] = value;
+        stack[0] = pos;
+    }
 
-        public void push(long value) {
-            if (pos < cap) {
-                array[pos++] = value;
-            } else {
-                // if array is full, reset pos
-                pos = 0;
-                array[pos++] = value;
-            }
+    static long pop(long[] stack) {
+        int cap = stack.length - 1;
+        int pos = (int) stack[0];
+        if (pos > 0) {
+            long value = stack[pos];
+            stack[0] = pos - 1;
+            return value;
         }
 
-        public long pop() {
-            if (pos > 0) {
-                pos--;
-                return array[pos];
-            } else {
-                pos = cap;
-                pos--;
-                return array[pos];
-            }
-        }
+        pos = cap;
+        long value = stack[pos];
+        stack[0] = pos - 1;
+        return value;
     }
-}
\ No newline at end of file
+}
diff --git a/core/src/test/java/com/taobao/arthas/core/util/LongStackTest.java b/core/src/test/java/com/taobao/arthas/core/util/LongStackTest.java
index 009e0473850..9b6d14b6be9 100644
--- a/core/src/test/java/com/taobao/arthas/core/util/LongStackTest.java
+++ b/core/src/test/java/com/taobao/arthas/core/util/LongStackTest.java
@@ -3,8 +3,6 @@
 import org.junit.Assert;
 import org.junit.Test;
 
-import com.taobao.arthas.core.util.ThreadLocalWatch.LongStack;
-
 /**
  * 
  * @author hengyunabc 2019-11-20
@@ -14,76 +12,76 @@ public class LongStackTest {
 
     @Test
     public void test() {
-        LongStack stack = new LongStack(100);
+        long[] stack = new long[101];
 
-        stack.push(1);
-        stack.push(2);
-        stack.push(3);
+        ThreadLocalWatch.push(stack, 1);
+        ThreadLocalWatch.push(stack, 2);
+        ThreadLocalWatch.push(stack, 3);
 
-        Assert.assertEquals(3, stack.pop());
-        Assert.assertEquals(2, stack.pop());
-        Assert.assertEquals(1, stack.pop());
+        Assert.assertEquals(3, ThreadLocalWatch.pop(stack));
+        Assert.assertEquals(2, ThreadLocalWatch.pop(stack));
+        Assert.assertEquals(1, ThreadLocalWatch.pop(stack));
     }
 
     @Test
     public void test2() {
-        LongStack stack = new LongStack(100);
+        long[] stack = new long[101];
 
-        stack.push(1);
-        stack.push(2);
-        stack.push(3);
+        ThreadLocalWatch.push(stack, 1);
+        ThreadLocalWatch.push(stack, 2);
+        ThreadLocalWatch.push(stack, 3);
 
-        Assert.assertEquals(3, stack.pop());
-        Assert.assertEquals(2, stack.pop());
-        Assert.assertEquals(1, stack.pop());
-        Assert.assertEquals(0, stack.pop());
+        Assert.assertEquals(3, ThreadLocalWatch.pop(stack));
+        Assert.assertEquals(2, ThreadLocalWatch.pop(stack));
+        Assert.assertEquals(1, ThreadLocalWatch.pop(stack));
+        Assert.assertEquals(0, ThreadLocalWatch.pop(stack));
     }
 
     @Test
     public void test3() {
-        LongStack stack = new LongStack(2);
+        long[] stack = new long[3];
 
-        stack.push(1);
-        stack.push(2);
-        stack.push(3);
+        ThreadLocalWatch.push(stack, 1);
+        ThreadLocalWatch.push(stack, 2);
+        ThreadLocalWatch.push(stack, 3);
 
-        Assert.assertEquals(3, stack.pop());
-        Assert.assertEquals(2, stack.pop());
-        Assert.assertEquals(3, stack.pop());
-        Assert.assertEquals(2, stack.pop());
+        Assert.assertEquals(3, ThreadLocalWatch.pop(stack));
+        Assert.assertEquals(2, ThreadLocalWatch.pop(stack));
+        Assert.assertEquals(3, ThreadLocalWatch.pop(stack));
+        Assert.assertEquals(2, ThreadLocalWatch.pop(stack));
     }
 
     @Test
     public void test4() {
-        LongStack stack = new LongStack(2);
+        long[] stack = new long[3];
 
-        stack.push(1);
-        stack.push(2);
+        ThreadLocalWatch.push(stack, 1);
+        ThreadLocalWatch.push(stack, 2);
 
-        Assert.assertEquals(2, stack.pop());
-        Assert.assertEquals(1, stack.pop());
-        Assert.assertEquals(2, stack.pop());
-        Assert.assertEquals(1, stack.pop());
+        Assert.assertEquals(2, ThreadLocalWatch.pop(stack));
+        Assert.assertEquals(1, ThreadLocalWatch.pop(stack));
+        Assert.assertEquals(2, ThreadLocalWatch.pop(stack));
+        Assert.assertEquals(1, ThreadLocalWatch.pop(stack));
     }
     
     @Test
     public void test5() {
-        LongStack stack = new LongStack(10);
-
-        stack.push(1);
-        stack.push(2);
-        stack.push(3);
-        stack.pop();
-        stack.pop();
-        stack.push(4);
-        stack.push(5);
+        long[] stack = new long[11];
+
+        ThreadLocalWatch.push(stack, 1);
+        ThreadLocalWatch.push(stack, 2);
+        ThreadLocalWatch.push(stack, 3);
+        ThreadLocalWatch.pop(stack);
+        ThreadLocalWatch.pop(stack);
+        ThreadLocalWatch.push(stack, 4);
+        ThreadLocalWatch.push(stack, 5);
         
-        stack.push(6);
-        stack.pop();
+        ThreadLocalWatch.push(stack, 6);
+        ThreadLocalWatch.pop(stack);
 
-        Assert.assertEquals(5, stack.pop());
-        Assert.assertEquals(4, stack.pop());
-        Assert.assertEquals(1, stack.pop());
-        Assert.assertEquals(0, stack.pop());
+        Assert.assertEquals(5, ThreadLocalWatch.pop(stack));
+        Assert.assertEquals(4, ThreadLocalWatch.pop(stack));
+        Assert.assertEquals(1, ThreadLocalWatch.pop(stack));
+        Assert.assertEquals(0, ThreadLocalWatch.pop(stack));
     }
 }
diff --git a/pom.xml b/pom.xml
index 6d4d8c99159..6511a8cdb7b 100644
--- a/pom.xml
+++ b/pom.xml
@@ -109,7 +109,7 @@
             <dependency>
                 <groupId>com.alibaba.middleware</groupId>
                 <artifactId>termd-core</artifactId>
-                <version>1.1.7.14</version>
+                <version>1.1.7.15</version>
             </dependency>
             <dependency>
                 <groupId>com.alibaba.middleware</groupId>