Researchers introduced MemPrivacy, an edge-cloud framework designed to protect user data in LLM-powered agents while preserving memory utility. The system functions by detecting privacy-sensitive information on edge devices and replacing it with semantically structured type-aware placeholders before sending data to the cloud. Original values are stored locally in a secure SQLite database and restored to responses after cloud processing, ensuring that sensitive data like credentials, medical records, and PII (personally identifiable information) remain isolated from cloud-side memory systems.

The framework introduces a four-level privacy taxonomy (PL1–PL4) to support configurable protection policies, ranging from low-sensitivity preferences to critical secrets like recovery codes and API keys. MemPrivacy was evaluated using MemPrivacy-Bench, a new dataset featuring 200 synthetic users and over 52k privacy instances across multi-turn dialogues. Testing shows that MemPrivacy-4B-RL achieves F1 scores of 85.97% and 94.48% in privacy information extraction, outperforming general-purpose models like GPT-5.2 and Gemini-3.1-Pro.

Evaluation across existing memory systems such as Mem0, LangMem, and Memobase demonstrates that MemPrivacy effectively balances privacy and utility. When applying stringent protection policies (PL2–PL4), system accuracy loss is limited to 1.6%, significantly better than traditional masking methods that often destroy task semantics. The framework is designed for efficient on-device deployment, maintaining processing latency per message below one second. The research team released multiple model variants including 4B and 1.7B parameter versions trained via SFT (supervised fine-tuning) and RL (reinforcement learning) methods.

Researchers introduced MemPrivacy, an edge-cloud framework designed to protect user data in LLM-powered agents while preserving memory utility. The system functions by detecting privacy-sensitive information on edge devices and replacing it with semantically structured type-aware placeholders before sending data to the cloud. Original values are stored locally in a secure SQLite database and restored to responses after cloud processing, ensuring that sensitive data like credentials, medical records, and PII (personally identifiable information) remain isolated from cloud-side memory systems.

The framework introduces a four-level privacy taxonomy (PL1–PL4) to support configurable protection policies, ranging from low-sensitivity preferences to critical secrets like recovery codes and API keys. MemPrivacy was evaluated using MemPrivacy-Bench, a new dataset featuring 200 synthetic users and over 52k privacy instances across multi-turn dialogues. Testing shows that MemPrivacy-4B-RL achieves F1 scores of 85.97% and 94.48% in privacy information extraction, outperforming general-purpose models like GPT-5.2 and Gemini-3.1-Pro.

Evaluation across existing memory systems such as Mem0, LangMem, and Memobase demonstrates that MemPrivacy effectively balances privacy and utility. When applying stringent protection policies (PL2–PL4), system accuracy loss is limited to 1.6%, significantly better than traditional masking methods that often destroy task semantics. The framework is designed for efficient on-device deployment, maintaining processing latency per message below one second. The research team released multiple model variants including 4B and 1.7B parameter versions trained via SFT (supervised fine-tuning) and RL (reinforcement learning) methods.