SOLAR: learning-augmented semantic cache replacement outperforms FIFO

Researchers propose SOLAR, a learning-augmented framework for online semantic cache replacement that addresses the failure of classic heuristics like LRU and LFU on workloads lacking temporal locality. The method derives modification timing from regret accumulation and content selection from Bayesian online learning over implicit retrieval feedback.

• Experiments on MemoryBench-Full datasets (LoCoMo, DialSim) show SOLAR achieves 5–75% relative improvement over FIFO at tight cache sizes.
• The framework proves a constant competitive ratio ≤ 3, independent of cache size and horizon, compared to Ω(K) for FIFO.
• Synthetic experiments with 5000-item pools reveal an inverted-U relationship between pool size and retrieval quality.

The authors consider this important because it justifies capacity constraints as a retrieval noise phenomenon rather than a storage limitation.