One of the challenges in using intra-cortical recordings like Local Field Potentials for Brain Computer Interface (BCI) is their inherent day-to-day variability and non-stationarity caused by subject motivation and learning. Practical Brain Computer Interfaces need to overcome these variations, as models trained on characteristic features from one day fail to represent new characteristics of another. This paper proposes a novel adaptive model that adjusts to signal variation by appending new features to the existing model and without knowledge of actual hand kinetics in an unsupervised way. With this adapting model we investigated the effects of learning and model adaptation on BCI performance. Using this new model we dramatically improve on all previously published long term decoding and show that target direction is accurately decoded in 95% of the trials over two weeks and in 85% of the trials in varying environments. Since the model needs no separate re-calibration, it can reduce user frustration and improve BCI experience.