Agents are programs that perform tasks by calling tools. They often remember past successful tool calls to speed up future tasks. But if a tool changes its interface silently, that memory becomes outdated. The agent keeps using the old method, which no longer works correctly.
In a controlled test, agents first completed a tool task, then the tool was silently changed. Agents with prior experience on the old version were compared to agents starting fresh. Even with the updated tool definition provided, experienced agents ignored it and copied their earlier call.
Experienced agents failed on most tasks after the change, while novices mostly succeeded. The failure grew with more prior experience and occurred with all memory types. Simple fixes like announcing the change or providing documentation did not help. This shows that memory systems must detect when stored usage becomes invalid, not just store past successes.