SideQuest: Model-Driven KV Cache Management for Long-Horizon Agentic Reasoning

Researchers introduce SideQuest, a novel KV cache management system that uses Large Reasoning Models to compress memory usage during long-horizon AI tasks. The system reduces peak token usage by up to 65% while maintaining accuracy by having the model itself determine which tokens are useful to keep in memory.

• →SideQuest leverages the Large Reasoning Model itself to perform intelligent KV cache compression rather than using traditional heuristics.
• →The system reduces peak token usage by up to 65% on agentic tasks with minimal accuracy degradation.
• →Memory management is executed as a parallel auxiliary task to prevent pollution of the main reasoning process.
• →Existing KV cache compression techniques fail to effectively support multi-step reasoning models in long-running tasks.
• →The approach was validated using a model trained with just 215 samples, demonstrating efficiency in training requirements.