Abstract
We present Category-Adaptive Recall Depth (CARD), a retrieval mechanism for heat-tiered memory systems that conditions top-K depth and ranking strategy on a query category signal. Prior memory retrieval systems apply a fixed top-K regardless of query semantics; CARD instead assigns each query to one of four categories—information extraction (IE), knowledge update (KU), multi-session reasoning (MSR), temporal reasoning (TR)—and selects a per-category (K, ranking_strategy) pair optimized for that category's retrieval demands. KU queries additionally filter to the highest-version record per entity, enabling 100.0% accuracy on knowledge update tasks. We evaluate CARD integrated with LOCI (Layered Output and Context Intelligence), a heat-tiered memory system with versioned fact storage, against LongMemEval-500 (Zhang et al., ICLR 2025), a 500-question benchmark across the four query categories. Using Microsoft phi4-14B on a consumer NVIDIA RTX 3080 Ti at zero marginal API cost, CARD+LOCI achieves: IE=94.2%, KU=100.0%, MSR=97.7%, TR=96.2%, overall=96.6%. This result is approximately 14-15 percentage points above the reported GPT-4 baseline (~82%) on the same benchmark. We disclose the CARD mechanism, per-category configurations, and full benchmark methodology as defensive prior art to prevent enclosure of category-adaptive retrieval strategies in AI memory systems.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Sanyal, Arko, "Category-Adaptive Recall Depth (CARD) for Heat-Tiered Memory Systems, with Experimental Validation on LongMemEval-500", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/10856
Raw benchmark results: CARD+LOCI+phi4-14B on LongMemEval-500. n=500, overall=96.6%, KU=100.0%, IE=94.2%, MSR=97.7%, TR=96.2%. Generated 2026-07-10 on RTX 3080 Ti.