diff --git a/kcws/train/train_cws.py b/kcws/train/train_cws.py
index e68cdeb..1dad7a1 100644
--- a/kcws/train/train_cws.py
+++ b/kcws/train/train_cws.py
@@ -32,6 +32,7 @@
 tf.app.flags.DEFINE_integer("train_steps", 150000, "trainning steps")
 tf.app.flags.DEFINE_float("learning_rate", 0.001, "learning rate")
 tf.app.flags.DEFINE_bool("use_idcnn", True, "whether use the idcnn")
+tf.app.flags.DEFINE_integer("track_history", 6, "track max history accuracy")
 
 
 def do_load_data(path):
@@ -200,6 +201,7 @@ def test_evaluate(sess, unary_score, test_sequence_length, transMatrix, inp,
             total_labels += sequence_length_
     accuracy = 100.0 * correct_labels / float(total_labels)
     print("Accuracy: %.3f%%" % accuracy)
+    return accuracy
 
 
 def inputs(path):
@@ -232,6 +234,8 @@ def main(unused_argv):
         with sv.managed_session(master='') as sess:
             # actual training loop
             training_steps = FLAGS.train_steps
+            trackHist = 0
+            bestAcc = 0
             for step in range(training_steps):
                 if sv.should_stop():
                     break
@@ -244,9 +248,22 @@ def main(unused_argv):
                         print("[%d] loss: [%r]" %
                               (step + 1, sess.run(total_loss)))
                     if (step + 1) % 1000 == 0 or step == 0:
-                        test_evaluate(sess, test_unary_score,
-                                      test_sequence_length, trainsMatrix,
-                                      model.inp, tX, tY)
+                        acc = test_evaluate(sess, test_unary_score,
+                                            test_sequence_length, trainsMatrix,
+                                            model.inp, tX, tY)
+                        if acc > bestAcc:
+                            if step:
+                                sv.saver.save(
+                                    sess, FLAGS.log_dir + '/best_model')
+                            bestAcc = acc
+                            trackHist = 0
+                        elif trackHist > FLAGS.track_history:
+                            print(
+                                "always not good enough in last %d histories, best accuracy:%.3f"
+                                % (trackHist, bestAcc))
+                            break
+                        else:
+                            trackHist += 1
                 except KeyboardInterrupt, e:
                     sv.saver.save(sess,
                                   FLAGS.log_dir + '/model',