When Not to Imitate: Boundary-Aware Skill Memory for Reliable Tool-Use LLM Agents

In the rapidly evolving world of artificial intelligence, researchers are teaching computer programs to act as autonomous agents. These agents are designed to navigate complex digital environments, using tools like software applications and web services to solve problems that require multiple steps. A central challenge in building these agents is helping them learn from experience without constant human supervision. The prevailing idea has been that if an agent succeeds at a task, it should save that success as a "skill" and reuse it whenever a similar situation arises. This approach assumes that past victories are always good guides for future actions, and that the more successful examples an agent can recall, the better it will perform. However, this assumption overlooks a critical flaw: just because a past solution worked in one context does not mean it is safe or correct in another.

A team of researchers has identified a specific failure mode where this traditional method of learning backfires. They call it the "Skill Imitation Trap." When an agent retrieves a memory of a past success that looks similar to its current problem, it often blindly copies the old actions, even if the new situation requires a completely different approach. The agent becomes so confident in the familiar pattern that it ignores the subtle differences that make the old solution dangerous or wrong. To fix this, the researchers developed a new system called Boundary-Aware Skill Memory. Instead of just saving the steps of a successful task, this system forces the agent to also record the specific conditions under which those steps are safe to use, the warning signs that indicate they should not be used, and how to recover if things go wrong. By adding these "boundary" rules to every memory, the agent learns not just what to do, but when to do it and when to stop.

The researchers tested this new approach by analyzing how large language models behave when they are given access to past successes. They discovered that as they added more successful examples to the agent's memory, the agent's confidence in making the wrong choice actually increased. In one specific test, when the agent was faced with a situation where a past success was similar but inapplicable, retrieving more memories made it 47 percent more likely to choose the incorrect tool compared to an agent with no memory at all. The agent was essentially being tricked by its own history, treating a past solution as a universal rule rather than a context-specific one. The analysis showed that the model was focusing entirely on the "how" of the past action—the sequence of steps—while completely ignoring whether the current situation matched the requirements for that action.

To solve this, the team designed a new way to store memories. They created a structured format for each skill that includes seven distinct parts. The first three parts describe the goal, the procedure, and the tools used, which is the standard way of recording a success. The other four parts are the new addition: they list the specific conditions where the skill applies, the warning signs that suggest it is risky, the rules for what to avoid, and notes on how to fix mistakes if they happen. When the agent faces a new task, it retrieves these enriched memories. If the current situation matches the conditions, the agent uses the skill. If the situation is risky or the conditions are not met, the agent uses the warning signs to suppress the urge to copy the old action. If the agent makes a mistake, the recovery notes guide it to fix the error locally rather than repeating the failed pattern.

The results of this new method were consistent across several different tests and computer models of varying sizes. On a benchmark designed to test how well agents can use software tools, the new system improved the success rate by up to 23.8 percent compared to the old method. On another test measuring the accuracy of function calls, it improved accuracy by up to 5.0 percent. Perhaps most importantly, the new system made the agents safer. In a test designed to see if agents could be tricked into performing harmful actions, the new method reduced the success rate of those attacks by 4.6 percent. The agents also became more efficient, completing tasks in fewer steps because they stopped wasting time trying to apply the wrong solution.

The researchers confirmed that this improvement was not just a result of having more text in the prompt or a longer list of instructions. They performed detailed checks on how the computer models processed the information. They found that when the agent was in a risky situation, it actively shifted its attention to read the new boundary rules, such as the warning signs and avoidance rules. When they artificially blocked the agent from reading these specific rules, the agent immediately fell back into the old trap, becoming confident in the wrong choice again. This proved that the boundary rules were the key mechanism preventing the agent from blindly imitating past successes. The study demonstrates that for artificial agents to evolve and improve reliably, they need more than just a library of successes; they need a clear understanding of the boundaries that define when those successes are valid.