Alibaba Cloud Launches Memory Service on Model Studio, Claims Top Scores on HaluMem and PersonaMem-v2

Alibaba Cloud has launched Memory, a long-term memory service for personal AI agents, on its Model Studio platform. IT Home reported the launch on November 2. The service lets agents remember a user's preferences, plans and earlier commitments across conversations and assembles that history into a structured user profile the agent can draw on later.

Benchmark claims

Alibaba Cloud says the system leads two public memory benchmarks. In charts released with the launch, the system appears under the name Mementis.

• HaluMem-medium: it ranks first on four factual-memory metrics, including recall, precision and F1. Its question-answering score is 85.15%, against 81.71% for Zep, the next system listed.
• PersonaMem-v2: it posts an overall score of 56.82, first among eight systems compared and highest on seven of nine tasks. MemBrain is second at 55.72.

Alibaba Cloud notes that results for the other systems are taken from the OmniMemEval evaluation. The figures come from the company and have not been independently reproduced.

Developer controls and integration

Developers can define their own memory rules and profile fields to decide what an agent should keep. The console provides debugging tools and change tracking so teams can see how a memory was formed or updated.

Memory can be called through an API, or connected through MCP and a command-line tool to agent environments including Codex, DeepSeek Harness, Claude Code and OpenClaw. It also works with Skills, letting agents combine stored context with packaged capabilities. Pricing not yet disclosed.

Example use case

Alibaba Cloud gives the example of a wealth advisory agent. It can recall a client's risk appetite, goals and past discussions, so later recommendations stay consistent with what the user has already said rather than starting from scratch in each session.

Strategic context

The launch fits into Alibaba's wider push on personal agents. At CNCC last month, Qwen's Jiang Guanjun described a personal agent framework built on three pillars: models, memory and an ecosystem of capabilities.

What this means

Memory is offered as a separate managed service rather than something each agent team builds itself. Custom memory rules, editable profile fields and change tracking give developers direct control over what an agent retains and a way to inspect how a stored memory came to be. Support for MCP and a command-line tool means the service can be attached to several existing agent environments, not only Alibaba's own.

The benchmark results are Alibaba Cloud's own and are unreproduced. For the comparison systems, the company says the scores come from the OmniMemEval evaluation. Teams evaluating the service should treat the HaluMem-medium and PersonaMem-v2 rankings as claims to test against their own workloads. The launch also matches the framework Jiang Guanjun outlined at CNCC, which names memory as one of three pillars of a personal agent alongside models and an ecosystem of capabilities.