CrystalCache: Cross-Domain Transfer from Cognitive Memory Crystallization to KV Cache Eviction in Long-Context LLMs

We present CrystalCache, a KV-cache eviction algorithm for long-context large language models, derived by porting a cognitive theory of memory crystallization into the cache-management domain. CrystalCache addresses memory bottlenecks by organizing past context into semantic units ("trunks") and scoring each unit along two independent dimensions: associative crystallization (D), which captures how strongly a trunk has been re-engaged across queries, and encoding impact (M_i), which captures how distinctively that trunk has shaped past attention. Eviction combines these signals with an explicit rarity term and replaces binary keep/drop decisions with progressive retention, so high-value content degrades gracefully rather than being abruptly discarded. Across Llama-3.1-8B, Mistral-7B, and Qwen3-8B, CrystalCache yields consistent improvements on long-context retrieval benchmarks, with the most pronounced gains on Needle-in-a-Haystack tasks where rare-but-critical evidence is precisely the failure mode standard eviction policies handle worst.

• KV cache eviction
• long-context LLM
• memory crystallization
• cross-domain transfer
• associative scoring
• progressive retention
• Needle-in-a-Haystack