On November 2, Alibaba Cloud upgraded its Memory service to enhance long-term memory and personalization capabilities for Personal Agent scenarios, supporting MCP and CLI integration. It can extract conversational data in real time to generate user profiles, dynamically track preferences in financial advisory contexts, and deliver intelligent recommendations. In evaluations, it outperformed competitors on multiple metrics and has been seamlessly integrated with leading Agent tools such as Codex and DeepSeek Harness, advancing the evolution of personal assistants toward deeper personalization and long-term companionship.Author and source: AIBase
On November 2, Alibaba Cloud announced a comprehensive upgrade to its Memory service, focusing on enhancing long-term memory and personalized processing capabilities for user preferences, plans, and historical agreements in Personal Agent scenarios, while fully supporting mainstream integration methods such as [MCP] and [CLI].
This service extracts key information from conversations in real time and automatically saves it as user profiles and structured memories according to predefined rules. In subsequent interactions, the system can precisely retrieve and apply relevant context based on the user’s identity and current query, enabling continuous follow-up on user needs and effectively reducing repetitive questions.
Evaluation results show that, in the recent [HaluMem-medium] benchmark, Memory achieved the highest scores across recall, accuracy [F1], and the C metric for memory-based question answering. In the [PersonaMem-v2] evaluation, it ranked first among the [8] participating systems with an overall score of [56.82], securing the top position in [7] out of 9 task dimensions.
In real-world applications, such as financial advisory, the system can automatically summarize users’ risk preferences and liquidity needs to generate a profile, then collaborate with [Agent] in subsequent conversations to perform product comparisons and risk alerts, dynamically updating its memory when user preferences change. Currently, developers can customize memory rules, profile fields, and tracking and retrieval effectiveness via the console, and seamlessly integrate this capability into leading [Agent] tools and frameworks such as Codex, DeepSeek Harness, Claude Code, and OpenClaw via API, MCP, or CLI. Additionally, [Memory] can deeply cooperate with the [Skill] module to decouple individual preferences from reusable task workflows, enabling efficient collaboration.
This upgrade marks a significant step forward for the personal assistant as it breaks through the context limitations of traditional single-turn conversations, evolving toward deep personalization and long-term companionship, providing a solid infrastructure foundation for [Agent] to operate effectively in complex vertical scenarios.


