Most AI tools have a fundamental problem: they forget you the moment the conversation ends.
Every time you open a new chat, you’re a stranger again. You re-explain your context, your preferences, your project names. The AI is smart, but it has no memory of who you are. It’s like hiring a brilliant consultant who gets amnesia every morning.
I wanted something different. I wanted an AI that knows me.
What is Link?
Link is my personal AI collaborator, built on top of Claude and running inside CatDesk. It’s not a generic chatbot — it’s an AI that has been carefully configured with:
- My identity: who I am, what I do, how I think
- My work context: the Meituan product ecosystem, my current projects, my team’s goals
- My knowledge base: detailed notes on recommendation systems, marketing algorithms, search, and more
- My tools: connected to internal data platforms, document systems, and automation workflows
The key insight is that memory is designed, not discovered. You don’t wait for an AI to learn about you — you architect the knowledge it needs, layer by layer.
The architecture
Link uses a four-layer memory system:
L1 — Soul (link_soul.md): The core identity file. Who Link is, what it cares about, its working principles. Read every conversation.
L2 — Knowledge (core_knowledge.md): Five domains of deep expertise — subsidy algorithms, recommendation systems, marketing algorithms, search, and the NEXT platform. Read every conversation.
L3 — Status (current_status.md): What’s happening right now. Current tasks, recent decisions, work in progress. Read every conversation.
L4 — Details (knowledge/business/*.md): Deep-dive notes on specific topics. Read only when needed.
This layered approach keeps the context window clean. You don’t dump everything into every conversation — you load what’s relevant, when it’s relevant.
The tradeoffs
What works well:
- Zero re-explaining. Link knows my projects by name, knows my team’s terminology, knows what I care about.
- Domain-specific reasoning. When I ask about subsidy strategy, Link doesn’t give generic advice — it reasons within the specific constraints of our platform.
- Workflow automation. Skills for data queries, document generation, browser control — all accessible via natural language.
What’s still hard:
- Knowledge staleness. The files need to be updated manually. If I make a major decision and don’t update
current_status.md, Link won’t know. - Context window limits. Even with layered loading, very long conversations can push out important context.
- The “brilliant but junior” problem. Link is great at execution but still needs me to set direction. It doesn’t proactively surface insights I didn’t ask for.
What I learned
The biggest lesson: identity before capability. Before you configure what an AI can do, configure who it is. An AI with a clear identity makes better decisions, maintains more consistent tone, and is more useful as a collaborator.
The second lesson: memory is a product problem, not a technical problem. The hard part isn’t storing information — it’s deciding what to store, how to structure it, and when to load it. That’s pure product thinking.
I’m still iterating on this. But even in its current form, Link has fundamentally changed how I work. It’s the first AI tool that feels like a genuine collaborator rather than a very fast search engine.