Self-reinforcing errors and over-generalization

Two subtler integrity failures from the "Practical Guide" deserve their own mention, because they are where memory turns actively harmful:

• Self-reinforcing errors (confirmation loops). The system treats a memory as ground truth, but that memory is wrong, and then it interprets everything through that wrong belief. The author's real example is striking: his system decided a smart-home integration was faulty, so it ignored all data from those devices as erroneous, when in reality there were just a few dead batteries. The agent built a wrong belief and then defended it.
• Over-generalization. The agent learns a lesson in a narrow context and then applies it everywhere. A workaround for one customer, or a fix for one specific error, becomes a default pattern applied where it does not belong.