Moving Memory Retrieval Into the Agent Kernel Beats Mem0 by Up to 4 Points and Cuts Latency 15-61%

arXiv 2609.10144·medium signalXBlueskyLinkedInCopy link

arXiv 2609.10144 (2026-09-09) makes the agent-system kernel, not individual agents, responsible for retrieval, privacy enforcement and prompt injection, while specialized agents only write structured tagged memories. Implemented on AIOS and tested across GPT-4o, Llama-3.1:8B and Qwen-2.5:7B over 1,800 trials, it improves personalization scores by 2.4 to 4.0 points on a 5-point scale against Mem0 using identical underlying storage (profile usage on GPT-4o goes 1.05 to 4.69), every comparison significant at p < 10^-18. Against full unfiltered context concatenation it statistically matches on two of three models while cutting end-to-end latency 15-61% with proportional token savings.