State Corruption

Almost every computer program keeps track of where it currently stands. An online shop remembers what’s in the shopping cart. A chat program remembers what was discussed before. This entirety of remembered information is called the state. State Corruption occurs when this remembered state breaks or becomes falsified. The program often doesn’t crash. It simply keeps computing, just with false assumptions about the world.

The error nobody notices

A crash is annoying, but honest. You immediately see that something is wrong. A corrupted state is more dangerous because it stays invisible. The system keeps delivering results, and at first glance they look plausible.

An example from banking makes this clear. A transfer consists of two steps: debiting and crediting. If the system fails after the first step, the money has vanished. The database is then in a state that shouldn’t be possible in the real world. This is precisely why databases go to enormous lengths to ensure that such half-completed processes don’t persist permanently.

With AI systems, a second problem arises. Errors in the state propagate. If an assistant that carries out several work steps in sequence stores an incorrect piece of information early on, it builds all further steps upon it. A small error becomes a long chain of wrong decisions.

How a state breaks

A classic cause is concurrent access. Two parts of a program modify the same stored value at the same moment. Both read the old value, both write their result back, and one of the two changes is lost. Experts call this a race condition, a race for the same memory location.

Another cause is interrupted processes. Power gone, network gone, process terminated: the state freezes mid-change. Faulty input also counts here. If a system accepts data without validating it, the nonsense goes straight into memory.

With AI agents, meaning programs that plan and execute multiple steps autonomously, state corruption often occurs without any technical defect at all. The agent writes down an intermediate result that it derived incorrectly. Technically, everything is fine; content-wise, it isn’t. This must be distinguished from hallucination: in a hallucination, the model invents an answer in the moment, whereas with State Corruption, the false value is permanently stored in the system’s memory and gets used again and again.

From security vulnerabilities to chatbot memory

The term appears regularly in security advisories. Attackers deliberately try to falsify a program’s memory state in order to use it for their own purposes. Many well-known vulnerabilities in operating systems and browsers fall into this category. Vendors release security updates to address them.

In everyday life, the phenomenon appears in a more harmless form. An app displays a shopping cart that was emptied long ago. A saved game suddenly becomes unusable. A chat history insists that you live in a different city, because an incorrect detail ended up in long-term memory. The standard advice to restart the app targets exactly this: you discard the broken state and start fresh.

In the development of AI agents, this topic is particularly prominent right now. Companies are therefore building in checkpoints where the state is reconciled against reality. Others deliberately have agents start with a fresh, empty memory once a task is completed. Both approaches cost computing time, but they prevent a single error from persisting through hours of work.