Tencent’s AI Design Agent Is Ugly. Its Memory System Is the Real Story.

The milk tea app homepage came out looking like it was designed by someone who had never seen a color wheel. Gray, muddy, vaguely depressing — like a rainy Tuesday rendered in pixels.

I felt that familiar pang of secondhand embarrassment you get when an AI tool promises brilliance and delivers something you’d hide from your portfolio.

This is Miora, Tencent’s first dedicated design agent. Launched today. Positioned as a competitor to Lovart. Capable of brand design, video, e-commerce ads, interactive games, and web applications.

And right now, it can’t design its way out of a paper bag.

The gap between what AI promises and what AI produces isn’t shrinking. It’s just getting better marketed.

The test was simple: design a homepage for a milk tea app. Miora has an “enhance prompt” button for when you don’t know what to type — which, let’s be honest, is most of us most of the time. I clicked it. I sent the prompt. I waited.

The result looked like someone took a perfectly good design and ran it through a gray filter. Twice. The colors were dirty. The layout was generic. It had that unmistakable AI-generated smell — technically functional, aesthetically offensive.

For comparison, I fed the same prompt to Codex. Better. Noticeably better. Not great, but at least it didn’t make me want to look away.

Miora does have editing features — dialog editing, where you tell the AI what to change, and DOM editing, where you tweak parameters through a Figma-like panel. It has an inspiration gallery. It has built-in design “Skills” you can upload, publish, or create through conversation. All of this is fine. None of it is the story.

AI design tools have mastered the art of almost. And almost is the most frustrating word in a creative professional’s vocabulary.

The story is Memory.

Miora has a four-layer memory system. User learns your creative habits, aesthetic preferences, and frequently used commands. Workflow stores your debugged AI creation steps for instant reuse. Project organizes visual assets, prompts, and results under specific project contexts. Correction records your feedback and fixes so the AI learns not to repeat mistakes.

On paper, this is brilliant. An AI that actually remembers what you like, what you’ve fixed, what your project needs. Not a blank slate every morning — a partner that accumulates knowledge.

Memory is the difference between an assistant and a colleague. One starts from zero every day. The other remembers what you argued about yesterday.

But here’s the twist nobody is talking about.

Memory cuts both ways.

If your taste is good, Miora’s memory system will learn your good taste. It’ll internalize your corrections, adopt your preferences, and gradually produce work that feels less like generic AI slop and more like you.

But if your taste is mediocre? If your corrections are wrong? If the feedback loop rewards safe, boring choices? Miora will learn that too. It’ll optimize for your mediocrity, lock you into a loop of “good enough,” and slowly train you to accept less while convincing you it’s improving.

An AI that learns your taste is an AI that learns your limits. And most of us don’t know which ones we’re teaching.

Then there’s the question nobody wants to ask: what exactly is being stored? Your aesthetic preferences. Your workflow patterns. Your project assets. Your creative process, deconstructed and archived on Tencent’s servers. Every correction you make is a data point. Every preference is a profile.

The memory feature isn’t just a creative tool — it’s a data collection engine dressed in designer clothing.

Now, Tencent’s positioning is clear. WorkBuddy is your general AI workbench — code, presentations, data, meetings. Miora is your specialized creative studio — brand visuals, video, 3D, full UI cases. The official analogy: WorkBuddy’s design features are like your phone’s built-in photo filters. Miora is a DSLR plus Photoshop.

Nice analogy. But a DSLR in the hands of someone who doesn’t understand photography still produces bad photos. A professional camera doesn’t make you a professional. And right now, Miora is a professional tool producing amateur results.

Specialization without quality isn’t innovation. It’s just a more specific kind of disappointment.

Here’s the broader signal that matters. AI design tools are splitting into two tiers: general-purpose assistants bolted onto workbenches, and specialized agents that claim to understand creative workflows. Miora represents the second wave — tools that don’t just generate, but remember, learn, and integrate into how you actually work.

For designers and product managers, this is the real takeaway. The tools are coming for your workflow, not just your output. They want to sit inside your process, learn your habits, and become indispensable through accumulated context. The question isn’t whether Miora’s current designs are ugly — they absolutely are. The question is whether you’re comfortable handing over your creative memory to a system that’s still learning to see.

The AI doesn’t need to be better than you today. It just needs to remember everything you teach it — and you’re teaching it every single time you click “correct.”

Miora isn’t ready. The designs are muddy. The output is underwhelming. But the architecture — the memory, the specialization, the integration with WorkBuddy — points to exactly where this is heading.

And that direction should make you feel two things at once: excitement that AI might finally become a creative partner worth having, and unease that the partnership requires you to surrender your taste, your process, and your data to a machine that currently can’t pick a color scheme.

The promise is real. The execution isn’t. And the memory system means every interaction you have with this mediocre tool is shaping the better one that comes after.

Choose your corrections wisely.