ACL2 is a mathematical logic, programming language, and mechanical theorem prover based on the applicative subset of Common Lisp. It is an "industrial-strength" version of the NQTHM or Boyer/Moore theorem prover, and has been used for the formal verification of commercial microprocessors, the Java Virtual Machine, interesting algorithms, and so forth.
AI::GA implements a (hopefully) generalized genetic algorithm. It does this by using an array of allowed tokens as individuals. The user has to provide a fitness function. There, the actual representation is implemented. If you have a string of chars, you can simply join them. If you want to have real numbers, you should probably use a bitwise representation and calculate the real values in your fitness function.
AIBash is a project which aims to make Bash act more intelligently. It features typing error-correction and the ability to learn that certain file suffixes are associated with certain programs so that other programs are filtered out while pressing TAB. Other features are planned for the future.
The Artificial Knowledge Interface for Reasoning Applications (AKIRA) project aims to create a C++ development framework to build cognitive architectures and complex artificial intelligent agents featuring KQML, fuzzy logic, neural networks, fuzzy cognitive maps, and DIPRA. DIPRA is a distributed version of the BDI (Belief Desire Intention) goal oriented model.
ASpiReNN is a little C library (with Python bindings) which provides support for simple (leaky integrate-and-fire) spiking neural networks. It is primarily designed for highly recurrent networks, but it can also be used with multi-layer nets, though performance won't be the same. Though only Leaky integrate-and-fire (for the neurons) and Spike-Timing Dependent Plasticity (for learning rules) are currently implemented, adding new models shouldn't be too difficult.
The ATRACO Project is a prototype implementation of a trusted ambient ecology system that runs and manages activity spheres in an Ambient Intelligence Space. Activity spheres are realized by automatically discovering, selecting, and adapting smart devices (artefacts) existing in the space, according to user's preferences, customs, and activities. OWL ontologies are used for modeling user profile, devices, activities, and goal descriptions. Abstract plans are bound to specific devices, methods, and values through semantic matching.