diff --git a/source/Rbm.cpp b/source/Rbm.cpp
index 4f1db0c..0d404bc 100644
--- a/source/Rbm.cpp
+++ b/source/Rbm.cpp
@@ -139,7 +139,8 @@ void Rbm::train(const arma::mat& batch, IListener* pListener)
 	arma::mat momentum_bias_v(arma::zeros(1, m_bv.n_cols));
 	arma::mat momentum_bias_hv(arma::zeros(1, m_bhv.n_cols));
 	arma::mat penalty_weights = arma::zeros(m_whv.n_rows, m_whv.n_cols);
-	
+	arma::mat ctx = arma::zeros(1, numContext());
+
 	int trainingSizeRemain = batch.n_rows;
 
 	bool shouldAbort = false;
@@ -158,8 +159,23 @@ void Rbm::train(const arma::mat& batch, IListener* pListener)
 		arma::mat v_probs(miniBatchSizeActual, m_bv.n_cols);
 		arma::mat hid_states(miniBatchSizeActual, m_bhv.n_cols);
 #endif
-		arma::mat c_states(arma::zeros(miniBatchSizeActual, m_ctx.n_cols));
-		arma::mat miniBatch(arma::join_rows(miniBatch_v, c_states));
+		arma::mat miniBatch;
+		if (numContext() > 0)
+		{
+			arma::mat c_states(arma::zeros(miniBatchSizeActual, numContext()));
+
+			for (int i=1; i < miniBatchSizeActual; i++)
+			{
+				arma::mat h = prob(v_to_h(arma::join_rows(miniBatch_v.row(i-1), ctx)));
+				ctx = h;
+				c_states.row(i) = ctx;
+			}
+			miniBatch = (arma::join_rows(miniBatch_v, c_states));
+		}
+		else
+		{
+			miniBatch = miniBatch_v;
+		}
 		arma::mat v_states(miniBatch);
 
 		// Create hidden layer base on training data