MEMTIER: Tiered Memory Architecture for Long-Running AI Agents — +33pp Accuracy Over Flat-File Baselines

arXiv·medium signal

New arXiv paper (2605.03675) proposes MEMTIER, a three-tier memory architecture for autonomous agents that introduces episodic JSONL storage, five-signal weighted retrieval, attention-attributed cognitive weight updates, async consolidation daemon, and PPO-based retrieval policy adaptation. Achieves Acc=0.382 on LongMemEval-S with Qwen2.5-7B on 6GB GPU — a 33 percentage point improvement over full-context baselines. Directly relevant to anyone building long-running agents.