Nowadays, there is an increasing need to provide a safe and independent living for cognitively deficient population. Notably, we have to improve seniors' autonomy and their public spaces accessibility. Giving these observations, the aim of this paper is to provide a personalized adaptive assisting system for elderly. More precisely, this paper presents the specification and implementation of a self-organizing multi-agent system able to abstract the different distributed components involved in user's environment. This system is able to detect different possible situations that a user could face in his daily outdoors activities and propose accordingly appropriate actions. This system not only learns user's habits from its perceptions but also improves its recommendations thanks to feedbacks provided by stakeholders (family, doctors, ...) following a reinforcement learning reasoning. Finally, we present our system evaluation specially its learning capabilities through different scenarios that have been generated automatically.