Stateful Memory Compaction for Mitigating Context Rot in Long-Horizon Agents

ABSTRACT

Long-horizon autonomous agents suffer from 'context rot,' a phenomenon where the accumulation of irrelevant interaction history severely degrades the model's reasoning capabilities over time. We propose Stateful Memory Compaction (SMC), a dynamic architecture that utilizes a secondary 'digestion' agent to continuously summarize, prune, and consolidate episodic memory into a dense semantic state vector. By enforcing a 'sawtooth' context length via periodic compaction, SMC maintains high reasoning fidelity across thousands of interaction turns. Empirical evaluations on multi-day coding and research tasks show that SMC reduces reasoning hallucinations by 64% while drastically lowering inference costs compared to infinite-context transformers.