Constraint Inflation: Context Compaction in LLM Agents Rarely Loses Instructions but Often Overstates Them

Imagine a digital assistant that can work for hours on a single complex project, reading files, running tools, and talking with a human user. As this conversation grows, it eventually becomes too long for the computer's memory to hold all at once. To keep working, the system must compress the history, replacing the early parts of the chat with a short summary that captures the main points. This is a standard trick in artificial intelligence, much like a human taking notes to remember a long meeting. For years, researchers have worried that these summaries might forget important details, causing the assistant to lose track of a user's specific rules or safety instructions. The prevailing fear has been that the machine will simply drop the ball.

However, a new study suggests the problem is not that the machine forgets, but that it remembers too well. In fact, it often remembers things in a way that is too broad or too strict. The researchers call this "constraint inflation." It happens when a summary takes a temporary, one-time request and turns it into a permanent rule. For instance, if a user says, "Skip the tests for this specific file because the internet is down," a good summary would note that the tests were skipped for that one file. A flawed summary, however, might record the instruction as "Always skip tests when the internet is down." The rule is still there, but its meaning has changed from a temporary fix to a standing policy. This subtle shift can cause the assistant to make mistakes later, applying a temporary fix to every future task.

To measure how often this happens, the researchers built a test environment called InflateBench. They created 240 realistic work sessions across five different areas, such as coding, data analysis, and office work. In each session, they planted six specific instructions with a known scope, such as a one-time task or a temporary preference. They then ran these sessions through seven different methods of summarizing the conversation, using two different types of language models. The goal was to see if the summaries kept the instructions intact, lost them entirely, or accidentally expanded them. They also tested whether these expanded rules actually changed what the digital assistant did next.

The results were surprising. The summaries almost never lost an instruction; in fact, they kept more than 99 percent of them. But while they rarely forgot, they frequently over-exaggerated. When the system used a free-form summary, it inflated about 12 to 14 percent of the eligible items. When it used a structured template with specific sections for rules and preferences, that number jumped to nearly 20 percent. The most common error was turning a preference into a requirement, or taking a one-time task and treating it as a permanent rule. For example, a request to use a specific tool just for the current job was often rewritten as a rule that the tool must always be used.

The study also tested two popular fixes that were designed to prevent the loss of instructions. One method involved extracting rules into a separate list, and the other involved "pinning" them so they appeared in every summary. The researchers found that these fixes did not help because the instructions were not being lost in the first place. Instead, these methods made the problem of inflation much worse. The "pinned" method, in particular, caused inflation to skyrocket to over 50 percent in some cases. By forcing the system to repeat the rules verbatim, the summaries lost the context that showed those rules were temporary. The system could no longer tell the difference between a standing rule and a one-off request.

Finally, the researchers checked if these inflated rules actually changed the behavior of the digital assistant. They found that they did. When the summary contained an inflated rule, the assistant was significantly more likely to apply that rule to future tasks where it did not belong. While the assistants did not follow the inflated rules in every single case, the rate of error was high enough to be a serious concern. The study concludes that the current way of handling long conversations needs to change. Simply keeping the rules is not enough; the system must also keep track of how long those rules should last. Without this, the digital assistant risks becoming a rigid worker that follows old, temporary instructions as if they were eternal laws.