From Cache to Cognition: A Grounded Cross-Domain Synthesis of KV Cache Management and Cognitive Memory Frameworks in LLM Inference

Version 2 — revised in response to an external structural review and an automated critique pass. See "Response to Review" appendix in the PDF for the change log.

The key-value (KV) cache has become a primary memory bottleneck in long-context large language model (LLM) inference, prompting a wave of compression and eviction strategies. Separately, cognitive and neuroscientific frameworks—including associative memory models and working-memory analogies—have been invoked to interpret how LLMs store and retrieve factual knowledge and how they reason. This paper offers a **heuristic cross-domain reading** of these two lines of work: not a formal derivation from shared structure, but an identification of where functional parallels are defensible given the cited sources, where they break down, and what testable questions the juxtaposition raises. We examine four concretely documented systems: TrimKV/DBTrimKV (XKV), IceCache, Reasoning in Memory (RiM), and Memory-Keyed Attention (MKA), alongside associative-memory and Hopfieldian analyses of LLM fact-learning. We argue that (1) globally calibrated KV eviction implements a form of utility-ranked selective retention whose functional role is *analogous to—but not derived from—*working-memory consolidation; (2) RiM's decoupling of internal computation from token generation mirrors the working-memory principle of not externalising every internal state, and this decoupling *may* reduce KV cache pressure as a secondary effect (unverified in the reviewed sources); and (3) hallucination is attributable to learning-phase failures rather than to KV eviction, despite surface similarity to source-monitoring errors. All claims are hedged to what the cited preprints—all unpublished, none peer-reviewed—directly support. ---

Authorship: Saluca Agentic AI Research Team (Saluca LLC). AI-drafted from arXiv preprint corpus on the date in the filename.

Cited arXiv preprints: 2605.30343v1