diff --git a/eo/src/eo b/eo/src/eo
index fd660b7b9..b87ca57f1 100644
--- a/eo/src/eo
+++ b/eo/src/eo
@@ -78,6 +78,7 @@
 #include <eoEvalCounterThrowException.h>
 #include <eoEvalTimeThrowException.h>
 #include <eoEvalUserTimeThrowException.h>
+#include <eoEvalKeepBest.h>
 
 // Continuators - all include eoContinue.h
 #include <eoCombinedContinue.h>
diff --git a/eo/src/eoEvalKeepBest.h b/eo/src/eoEvalKeepBest.h
new file mode 100644
index 000000000..0287cf1de
--- /dev/null
+++ b/eo/src/eoEvalKeepBest.h
@@ -0,0 +1,107 @@
+
+/*
+   This library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation;
+   version 2 of the License.
+
+   This library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with this library; if not, write to the Free Software
+   Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
+
+   © 2012 Thales group
+
+    Authors:
+        Johann Dreo <johann.dreo@thalesgroup.com>
+*/
+
+#ifndef eoEvalKeepBest_H
+#define eoEvalKeepBest_H
+
+#include <eoEvalFunc.h>
+#include <utils/eoParam.h>
+
+/**
+  Evaluate with the given evaluator and keep the best individual found so far.
+
+  This is useful if you use a non-monotonic algorithm, such as CMA-ES, where the
+  population's best fitness can decrease between two generations. This is
+  sometime necessary and one can't use elitist replacors, as one do not want to
+  introduce a bias in the population.
+
+  The eoEvalBestKeep is a wrapper around a classical evaluator, that keep the
+  best individual it has found since its instanciation.
+
+  To get the best individual, you have to call best_element() on the
+  eoEvalKeepBest itself, and not on the population (or else you would get the
+  best individual found at the last generation).
+
+  Example:
+
+    MyEval true_eval;
+    eoEvalKeepBest<T> wrapped_eval( true_eval );
+
+    // as an interesting side effect, you will get the best individual since
+    // initalization.
+    eoPop<T> pop( my_init );
+    eoPopLoopEval<T> loop_eval( wrapped_eval );
+
+    loop_eval( pop );
+
+    eoEasyEA algo( …, wrapped_eval, … );
+
+    algo(pop);
+
+    // do not use pop.best_element()!
+    std::cout << wrapped_eval.best_element() << std::endl;
+
+  @ingroup Evaluation
+  */
+template<class EOT> class eoEvalKeepBest : public eoEvalFunc<EOT>, public eoValueParam<EOT>
+{
+    public :
+        eoEvalKeepBest(eoEvalFunc<EOT>& _func, std::string _name = "VeryBest. ")
+            : eoValueParam<EOT>(EOT(), _name), func(_func) {}
+
+        virtual void operator()(EOT& sol)
+        {
+            if( sol.invalid() ) {
+                func(sol); // evaluate
+
+                // if there is no best kept
+                if( this->value().invalid() ) {
+                    // take the first individual as best
+                    this->value() = sol;
+                } else {
+                    // if sol is better than the kept individual
+                    if( sol.fitness() > this->value().fitness() ) {
+                        this->value() = sol;
+                    }
+                }
+            } // if invalid
+        }
+
+        //! Return the best individual found so far.
+        EOT best_element()
+        {
+            return this->value();
+        }
+
+        /** Reset the best individual to the given one. If no individual is
+         * provided, the next evaluated one will be taken as a reference.
+         */
+        void reset( const EOT& new_best = EOT() )
+        {
+            this->value() = new_best;
+        }
+
+    protected :
+        eoEvalFunc<EOT>& func;
+};
+
+#endif
diff --git a/eo/test/CMakeLists.txt b/eo/test/CMakeLists.txt
index 5f098d3bf..db415bebc 100644
--- a/eo/test/CMakeLists.txt
+++ b/eo/test/CMakeLists.txt
@@ -34,6 +34,7 @@ ENDIF()
 ######################################################################################
 
 SET (TEST_LIST
+  t-eoEvalKeepBest
   t-eoInitVariableLength
   t-eofitness
   t-eoRandom
diff --git a/eo/test/t-eoEvalKeepBest.cpp b/eo/test/t-eoEvalKeepBest.cpp
new file mode 100644
index 000000000..fdcdae866
--- /dev/null
+++ b/eo/test/t-eoEvalKeepBest.cpp
@@ -0,0 +1,78 @@
+#include <iostream>
+
+#include <es/make_real.h>
+#include <apply.h>
+#include <eoEvalKeepBest.h>
+#include "real_value.h"
+
+using namespace std;
+
+int main(int argc, char* argv[])
+{
+
+    typedef eoReal<eoMinimizingFitness> EOT;
+
+    eoParser parser(argc, argv);  // for user-parameter reading
+    eoState state;    // keeps all things allocated
+
+
+    /*********************************************
+     * problem or representation dependent stuff *
+     *********************************************/
+
+    // The evaluation fn - encapsulated into an eval counter for output
+    eoEvalFuncPtr<EOT, double, const std::vector<double>&>
+        main_eval( real_value ); // use a function defined in real_value.h
+
+    // wrap the evaluation function in a call counter
+    eoEvalFuncCounter<EOT> eval_counter(main_eval);
+
+    // the genotype - through a genotype initializer
+    eoRealInitBounded<EOT>& init = make_genotype(parser, state, EOT());
+
+    // Build the variation operator (any seq/prop construct)
+    eoGenOp<EOT>& op = make_op(parser, state, init);
+
+
+    /*********************************************
+     * Now the representation-independent things *
+     *********************************************/
+
+
+    // initialize the population - and evaluate
+    // yes, this is representation indepedent once you have an eoInit
+    eoPop<EOT>& pop   = make_pop(parser, state, init);
+
+    // stopping criteria
+    eoContinue<EOT> & term = make_continue(parser, state, eval_counter);
+
+    // things that are called at each generation
+    eoCheckPoint<EOT> & checkpoint = make_checkpoint(parser, state, eval_counter, term);
+
+    // wrap the evaluator in another one that will keep the best individual
+    // evaluated so far
+    eoEvalKeepBest<EOT> eval_keep( eval_counter );
+
+    // algorithm
+    eoAlgo<EOT>& ea = make_algo_scalar(parser, state, eval_keep, checkpoint, op);
+
+
+    /***************************************
+     * Now, call functors and DO something *
+     ***************************************/
+
+    // to be called AFTER all parameters have been read!
+    make_help(parser);
+
+    // evaluate intial population AFTER help and status in case it takes time
+    apply<EOT>(eval_keep, pop);
+
+    std::clog << "Best individual after initialization and " << eval_counter.value() << " evaluations" << std::endl;
+    std::cout << eval_keep.best_element() << std::endl;
+
+    ea(pop); // run the ea
+
+    std::cout << "Best individual after search and " << eval_counter.value() << " evaluations" << std::endl;
+    // you can also call value(), because it is an eoParam
+    std::cout << eval_keep.value() << std::endl;
+}