What they did

The authors built kernel-managed shared memory into AIOS. Specialized agents write tagged memories, while the agent-system kernel handles retrieval, privacy enforcement, and injection of selected memories into prompts.

They compared this design with an unmanaged Mem0 backend using the same underlying storage, standard retrieval-augmented injection, and full unfiltered context concatenation. The evaluation comprised 1,800 trials across GPT-4o, Llama-3.1:8B, and Qwen-2.5:7B.

Key findings

• Compared with unmanaged Mem0, kernel management raised personalization scores by 2.4–4.0 points on a five-point scale. GPT-4o's profile-usage score, for example, increased from 1.05 to 4.69.
• Every reported comparison with unmanaged memory was statistically significant at p < 10^-18; gains over standard retrieval-augmented injection were also large and consistent across all three models.
• The system statistically matched full, unfiltered context concatenation on two models and had a small, model-specific deficit on the third.
• Relative to full-context concatenation, kernel-managed injection reduced end-to-end latency by 15–61% across the three models, alongside lower token use and inference cost.

Why it matters

The results support treating shared memory as operating-system infrastructure rather than leaving each agent to retrieve and manage user context independently. This could make personalization available across specialized agents without repeatedly supplying an entire user history, while placing access and injection policies at a central enforcement point.

Caveats

The abstract does not describe the trial data, personalization rubric, memory scale, or privacy-threat evaluation, making it difficult to assess ecological validity and whether the kernel's privacy controls resist misuse. Results are limited to AIOS, three assistant models, and the tested baselines; the small deficit against full context on one model also suggests that effectiveness may depend on the underlying model.