Self-GC Manages Context for Long-Horizon LLM Agents.

Long-horizon tasks for Large Language Model (LLM) agents often lead to an accumulation of structured context, including tool results, files, plans, and user constraints. Current context management strategies, such as chronological pruning or self-summarization, are often insufficient. Heuristics can be blind to future dependencies, while summaries might obscure crucial evidence or editable artifacts. This paper introduces Self-GC, a novel framework that conceptualizes agent context not as disposable text, but as self-governing, indexed objects. Drawing an analogy to garbage collection, Self-GC doesn't just reclaim tokens; it actively manages the lifecycle of these context objects. A dedicated side-channel planner proposes actions like folding, masking, and pruning, which are then enforced by the system harness, ensuring recoverable sidecars and safe commit boundaries. Experimental results on challenging datasets demonstrate Self-GC's effectiveness. It pruned a significant percentage of prefix tokens while minimally impacting future continuations, outperforming heuristic baselines. In production, this led to a notable reduction in average input tokens, especially during peak usage. These findings suggest that robust context management for LLM agents requires runtime lifecycle control over structured, recoverable objects rather than simple text cleanup.