A memory leak was discovered in Mesa where model instances could never be garbage collected after agents were created. The root cause was the `Agent._ids` class attribute—a `defaultdict` that stored references to model instances to ensure `unique_id` values were unique on a per-model basis. Because `_ids` was a class-level attribute that persisted across the Python process, any model instance used as a key in this dictionary maintained a hard reference indefinitely, preventing the garbage collector from cleaning up the model and all its associated objects (agents, grids, etc.) even after the model went out of scope or was explicitly deleted.

This bug had significant practical consequences for Mesa users, particularly those running multiple simulations or batch experiments. Each time a model was instantiated and run within a function, the model objects would accumulate in RAM rather than being cleaned up when the function exited. This meant that running many model instances—common in parameter sweeps, sensitivity analyses, or optimization workflows—would cause unbounded memory growth, eventually exhausting available RAM. The issue was especially problematic because it was invisible to users: simply letting a model go out of scope or calling `del model` appeared to work but silently retained all the memory, and even explicitly removing agents with `model.remove_all_agents()` only partially addressed the problem depending on the space types used.

The fix moved the `unique_id` assignment logic from the `Agent` class into the `Model.register_agent()` method, eliminating the problematic class-level `_ids` defaultdict entirely. Instead of tracking IDs across all model instances in a shared dictionary, each model now maintains its own `agent_id_counter` instance attribute that starts at 1 and increments with each registered agent. This approach ensures that `unique_id` remains unique within each model instance while allowing the garbage collector to properly clean up model objects when they go out of scope, since there are no longer any persistent class-level references to model instances. The fix also replaced `itertools.count` with simple integer incrementation, which avoids upcoming pickle compatibility issues in Python 3.14.