Mycelial Cache: Topology-Aware Eviction Through Use-Dependent Structural Reinforcement

Classical cache eviction policies (LRU, LFU, ARC, W-TinyLFU) treat entries asindependent, scoring each by individual recency or frequency metrics. Real workloads exhibitco-access structure: entries accessed together form clusters, with certain entries serving asbridges connecting these clusters. We present the Mycelial Cache, a topology-aware cache thatmaintains a weighted mesh over entries where access patterns create weighted edges throughHebbian strengthening. Inspired by six convergent biological analogs across four phyla (fungalmycelial networks, neural synaptic plasticity, bone remodeling, vascular adaptation, slime moldtransport networks, and ant pheromone trails), the cache uses composite eviction scoring thatprotects structurally important entries. Deterministic replay tests show 83.7% hit rate versus82.6% (LRU) and 80.1% (moka) on 1K co-access cluster workloads, with 73.0% versus 66.2%(LRU) and 71.5% (moka) under a 1K workload shift. Criterion throughput at 100K showsa mixed cost profile: get() is 544.12 ns for Mycelial versus 12.13 ns for lru and 495.96 nsfor moka; set() is 1,174.25 ns versus 109.29 ns and 1,773.93 ns respectively. The currentimplementation exposes bridge-aware eviction, mesh introspection, and fever-based adaptation;predictive prefetching and topology-following invalidation remain future work.