Papers

Academic work, preprints, and technical reports.

arXiv preprint (arXiv:2608.20397) · 2026
21 min read
Nexus: Depth-Adaptive KV-Cache Splicing and Retrieval-Decoupled Tool Routing for Agentic LLMs on Unified Memory
Mustafa Arslan
Nexus decouples agentic tool routing from schema-prefill cost with an INT8 semantic lookaside buffer and a calibrated cross-encoder gate, and repairs relocated KV blocks with a depth-adaptive suffix redecode that never regresses below a full prefill.
arXiv preprint (arXiv:2601.15311v3) · 2026
22 min read
Aeon: High-Performance Neuro-Symbolic Memory Management for Long-Horizon LLM Agents
Mustafa Arslan
Aeon is a neuro-symbolic Cognitive Operating System that treats LLM agent memory as a managed OS resource: a SIMD-accelerated clustered vector index, an episodic graph, INT8 quantization, a decoupled write-ahead log, and a semantic lookaside buffer delivering sub-5µs retrieval.

Nexus: Depth-Adaptive KV-Cache Splicing and Retrieval-Decoupled Tool Routing for Agentic LLMs on Unified Memory