Alibaba Cloud launches Memory, a long-term memory service for personal AI agents
The Model Studio service builds a structured user profile from past conversations and plugs into Codex, Claude Code, DeepSeek Harness and OpenClaw via API, MCP or a command-line tool; Alibaba says it tops two memory benchmarks.
Why it matters
Memory that persists across sessions is one of the main gaps between chatbots and agents that act for people over weeks, and cloud providers are now selling it as a separate layer. Alibaba reports a HaluMem question-answering score of 85.15% against 81.71% for Zep, and 56.82 on PersonaMem-v2, first of eight systems; scores for rivals come from the OmniMemEval evaluation and none have been reproduced independently. It follows Tencent's open-source Hy Memory 2.0. Watch for pricing and independent tests.
Key facts
- Launched on Alibaba Cloud Model Studio
- HaluMem QA: 85.15% vs 81.71% for Zep
- PersonaMem-v2: 56.82, first of eight systems
- Connects via API, MCP or command-line tool
- Works with Codex, Claude Code, DeepSeek Harness and OpenClaw
- Benchmark figures are Alibaba's own