Alibaba Cloud launches Memory, a service that lets AI agents remember users across chats

What Alibaba launched

Alibaba Cloud has added a service called Memory to Model Studio, its platform for building AI applications. As pandaily.com reported, citing IT Home on November 2, it is a long-term memory service designed for personal AI agents, meaning assistants that act on behalf of one user.

The pitch is simple. An agent using Memory can remember a person's preferences, plans and earlier commitments from one conversation to the next. It then organizes that history into a structured user profile that the agent can consult later, instead of starting each session from a blank page.

How developers control what is kept

Memory is not an all-or-nothing recorder. Developers can define their own memory rules and profile fields, which lets them decide what an agent should hold on to and what it should ignore.

The console also includes debugging tools and change tracking. A team can look at how a particular memory was formed or updated, which matters when an agent says something about a user that looks wrong.

The company gives one example. A wealth advisory agent could recall a client's risk appetite, goals and past discussions. Later recommendations would then stay consistent with what the client has already said.

Where it plugs in

There are three ways in. Developers can call Memory through an API, which is a standard way for one program to request something from another. They can also use MCP, a connection standard for linking outside tools to agents, or a command-line tool.

Through those routes, Memory can connect to agent environments including Codex, DeepSeek Harness, Claude Code and OpenClaw. So it is not limited to agents built on Alibaba's own models.

It also works with Skills, which the source describes as packaged capabilities. An agent can combine stored context about the user with a ready-made ability, for example by applying what it knows about your preferences while running a packaged task.

The benchmark claims, and the caveat

Alibaba Cloud says its system leads two public tests of memory quality. In the charts released with the launch, it 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: its overall personalization score is 56.82, first among eight systems and highest on seven of nine tasks. MemBrain is second at 55.72.

Results for the other systems are taken from the OmniMemEval evaluation. The numbers come from Alibaba Cloud, and the source says they have not been independently reproduced. Treat them as a vendor's opening claim, not a settled ranking. Gaps of a few points on a benchmark may not show up in everyday use.

Part of a bigger personal-agent push

The launch fits a theme Alibaba has been pushing. Last month at CNCC, Qwen's Jiang Guanjun described a personal agent framework with three pillars: models, memory and an ecosystem of capabilities. Memory is the second of those pillars becoming a product.

The source also points to a similar move from a competitor. Tencent Hunyuan open-sourced Hy Memory 2.0, which adds working and code memory for agents. Memory is turning into a contested layer of agent software, not a side feature.

What the report does not say

The source covers features and benchmarks but does not mention pricing, regional availability or data-handling terms. That is a real gap for a service whose whole job is to store personal details about users.

Anyone considering it will want to know where memories are kept, who can read them, and how a user or developer deletes them. Those answers matter as much as the benchmark scores.