Alibaba Cloud upgrades Memory service for personal agent long-term memory
Alibaba Cloud upgraded its Memory service for personal agents, topping four HaluMem-medium metrics and scoring 56.82 on PersonaMem-v2 across eight systems.
Alibaba Cloud recently upgraded its Memory service, strengthening its ability to remember user preferences, plans and prior commitments for Personal Agent scenarios, and offering MCP and CLI access. Memory is one of the core services on Alibaba Cloud Bailian.
As agents move from one-off question answering toward continuous personal service, accurately remembering information, updating memory as needs change and calling it back in later tasks has become a key determinant of experience. Memory extracts key information from conversations and stores it as memories and user profiles according to rules, with agents retrieving relevant memories by user identity and current question in later sessions and adding them to task context, while newly valid information continues to be written back.
In HaluMem-medium comparisons Memory took the highest score on all four metrics covering recall, precision and F1 of memory extraction plus a memory question answering consistency indicator. On PersonaMem-v2 it placed first among the eight systems listed with an overall score of 56.82, taking the highest score in seven of nine task dimensions. HaluMem-medium mainly assesses extraction and use of historical information, while PersonaMem-v2 focuses on user understanding, implicit preference inference, preference updating and personalized response over long interactions.
For integration, developers can write, read and semantically retrieve memories through the API, or connect the capability into agent tooling and frameworks such as Codex, DeepSeek Harness, Claude Code and OpenClaw via MCP and CLI. In Bailian Agent Studio, developers can configure a memory store for an agent from the console or read and write memories through the memory plugin in workflow applications without building their own storage and retrieval pipeline.