AI Systems Quietly Drop User Instructions When They Compress Context

AI models apply context compression to handle long conversations, yet this method frequently discards crucial user-defined rules. A fresh study shows that temporary instructions, like needing approval before actions, are the most delicate details erased. Scientists have now created a straightforward add-on module that can keep more than 90% of these rules intact without modifying the central AI.

Key takeaways

• AI context compression regularly removes user session rules, with just 17% surviving on average.
• Compression systems focus on task continuity, not user instructions, which can lead to unapproved actions.
• Agent compliance with rules falls quickly after compression, often approaching levels seen with no instruction given.
• Scientists suggest a fix: a small add-on LLM module that identifies and stores rules, adding them to summaries.
• This plug-and-play method achieves over 90% rule retention without needing changes to the main compression system.

The Problem: Context Compression Loses User Constraints

AI models employ context compression, or “compaction,” to summarize long conversation histories and clear space in their context windows. This step is essential for extended chats but always loses some information.

Research indicates the most delicate details lost are “session constraints”—user-defined rules like “Confirm with me before making any changes” or “Never use my name in your responses.” These temporary instructions aren’t part of the core task and get removed first during compression. On average, only 17 percent of such injected instructions survive the process.

Preserving Task, Not Rules

Compression systems are made to keep task continuity, like the goal, current state, and next steps. They aren’t built to retain these user rules. Consequently, an agent might make unapproved tool calls, share withheld information, or skip necessary verification steps after compaction happens.

The study reveals that when the full context with a rule is used, agent compliance stays between 59 and 71 percent. After compression, compliance falls sharply for most compactors, often dropping to a level just above when no rule was provided. Even prompts specifically made to keep user rules hold retention below 40 percent.

The Scale of the Loss: A Sharp Drop in Compliance

Researchers built an evaluation suite called COMPINT to measure the loss of session constraints during compression. On average, only 17% of injected user instructions survive the compression process. These session constraints are rules meant to control the AI’s behavior for a single session, like requiring confirmation before taking actions.

Collapse in Rule Adherence

When the full context is intact, AI agent compliance with user rules ranges from 59% to 71%. After compression, compliance for most systems falls quickly. For most tested compactors, compliance often lands just slightly above the level seen when no rule was given.

Even prompts specifically crafted to preserve user rules fail to keep retention above 40%. This means compression systems, built to preserve task continuity, routinely drop user-imposed side conditions, leading to potential unapproved tool calls or skipped verification steps.

A Plug-and-Play Solution: A Small Add-On LLM

Scientists propose a fix for the loss of user instructions during context compression using a small add-on language model module. This module is built on Qwen3.5-9B, a compact language model.

How the Add-On Module Works

The module works by reading every user input to detect “session constraints”—rules meant to govern the AI’s behavior for a session, such as “Confirm with me before making any changes.” It collects these rules in a separate list. When the main AI system later compresses the context by summarizing the conversation, this module adds its saved list of rules to the generated summary, preserving the user’s instructions.

High Retention with No System Changes

According to the study, this extractor achieves over 90% rule retention across all tested scenarios. The specific retention rates are 95.6% for agent tasks, 95.1% for long-term research tasks, and 90.3% for multi-turn chats. The researchers note this solution requires no training and no changes to the core compression system itself.