The paper introduces kernel-managed shared memory, a system-level abstraction for multi-agent AI systems. In this design, agents write structured, tagged memories, but the agent-system kernel centrally governs retrieval, privacy enforcement, and . This is implemented on the AIOS framework and evaluated against three baselines: an unmanaged memory backend (Mem0), standard retrieval-augmented injection, and full, unfiltered context concatenation.
Across 1,800 trials with GPT-4o, Llama-3.1:8B, and Qwen-2.5:7B, kernel-managed retrieval and injection improved personalization scores by 2.4 to 4.0 points on a 5-point scale compared to the unmanaged backend, with all comparisons significant at p < 10^-18. Against full context, it statistically matched performance on two of three models and showed a small, model-specific deficit on the third, while using substantially shorter prompts. End-to-end latency was 15-61% lower across all models, with corresponding reductions in token usage and inference cost.
This work is interesting because it shows that centralizing memory management in the system kernel can deliver most of the personalization benefit of unconstrained context at a fraction of its cost, making system-wide personalization more practical for multi-agent settings.