Alibaba Qwen Names Three Pillars of Its Personal Agent: Model, Memory and Ecosystem

At the China Computer Conference 2026, held on 22 October, Jiang Guanjun, head of AI algorithms in Alibaba's Qwen division, laid out a technical approach to building a Personal Agent. His main thesis: an agent earns the word "personal" only when it creates incremental value and proactively serves a specific individual on its own. That is achievable only by combining three components: the model, memory and the ecosystem.

From Answering Questions to Taking Initiative

Jiang arranged the development of AI applications into stages. In his picture, Personal Agent sits on a new rung, where the very criterion of success changes. Previously, an application was judged on the quality of its answers: whether it explained something clearly, wrote a text correctly, or found information quickly. For a personal agent, that is not enough.

The talk named two requirements. First, the agent must create incremental value, meaning it does what the user would not have done alone, or would have taken noticeably longer to do. Second, the service must be personalized and proactive. The system does not wait for a command; it notices context and suggests the next step.

Proactivity is technically harder than reactive dialogue. It requires knowledge about the user, permission to act in external services, and enough reliability that the initiative does not become annoying.

Three Pillars Without Which the Agent Falls

By the logic of the talk, each of the three pillars covers its own part of the task:

• The model handles understanding intent, planning steps and calling tools. It determines whether the agent can break a vague request down into a sequence of actions. • Memory stores the user's preferences, history and context. Without it, personalization is impossible, and every conversation starts from zero. • The ecosystem gives the agent access to real services: shopping, payments, maps, booking. Without it, the agent can only advise, not act.

The wording "advance together" matters. Jiang does not propose improving the three components one after another. A strong model without memory produces generic answers, memory without a model becomes a warehouse of notes, and both without an ecosystem hit the boundary of the chat window.

Why Alibaba Has a Special Position Here

The third pillar is what most sets Alibaba apart from many competitors. The company has its own e-commerce, payment infrastructure through its affiliate Ant Group, and mapping and service products. For an agent, that is a ready-made set of actions it can perform inside a trusted environment, rather than through fragile integrations with third-party sites.

In China's consumer assistant market, this is a tangible advantage. Competitors such as ByteDance and Tencent are also building ecosystem bundles, so the contest is no longer about answer quality in benchmarks but about whose agent can do more useful things on the user's behalf. Speaking at a specialized academic venue suggests that Qwen wants to claim the technical framing of this debate as well.

What Teams Building Agents Should Check

For developers, the talk works as a checklist. First, memory should be designed as a separate subsystem: what to store, how data goes stale, and how the user can view and delete it. Second, proactivity requires a confidence threshold: an agent that acts too often quickly loses trust. Third, value should be measured not by conversation length but by time saved and tasks completed.

For business, the takeaway is simpler. Whoever owns the services the agent accesses, along with the data about the user, gains real leverage over it. The model is gradually becoming an interchangeable part, while memory and the ecosystem become what is hard to copy.