This week, Huolala (Lalamove's China brand) published a long-form post on Juejin, the Chinese developer community, tearing apart their self-built "memory system" for large models. What we find worth noting here isn't how flashy the tech is — it's that the article exposes a truth about enterprise AI: however powerful the model, get the memory layer wrong and your AI is a "goldfish brain."

What this is

Huolala (a same-city logistics company) runs an AI outbound system that cold-calls drivers. They hit a concrete problem: a driver says "no long-haul jobs" or "only take it if the price is higher" — next call, the AI starts from zero and runs the same script again.

The problem isn't model capability; it's engineering implementation: every time the model is invoked, the system only passes in the content of "this one call." What was said previously is nowhere in the context window.

The article draws a distinction worth highlighting: "history is not memory." History is a chronological dump of raw dialogue; memory is the structured, searchable, updatable extraction of facts worth keeping — preferences, objections, progress.

Huolala's solution splits in two:

  • Online side: after a session ends, the system automatically extracts key facts into the memory store; when the user asks something, hybrid retrieval (keyword + semantic similarity) surfaces relevant memories.
  • Offline side: the memory store is periodically "tidied up" — deduplication, retirement of stale entries, conflict resolution. The article even borrows the metaphor of human sleep (light sleep, deep sleep, REM sleep) to describe different layers of maintenance.

Industry view

The industry's take on building the "memory layer" splits into two camps.

The mainstream view treats it as one of the core bottlenecks of enterprise AI. Startups like LlamaIndex and Mem0 are betting hard on this direction; OpenAI's ChatGPT long-term memory and Anthropic's Projects feature are solving the same problem at the foundation-model level. An engineering practice at Huolala's scale validates the judgment that "off-the-shelf solutions aren't enough."

But the dissent deserves a hearing: these systems are brutally complex. The article itself documents extensive failure handling, retry logic, and concurrency control — a mid-sized company copy-pasting this architecture would almost certainly fail. A more realistic path, in our view, is to integrate a third-party memory middleware first, get the business running, then consider building in-house.

Another issue the article glosses over: memory systems make AI "understand" users better, but they also pile on privacy and compliance burden. Drivers' and users' conversations being long-term, structurally stored — where the compliance boundary sits, the piece doesn't say, but it's a question the industry can't dodge.

Impact on regular people

  • For enterprise IT: when evaluating internal AI projects, stop being dazzled by "how strong the model is." The model is just the tip of the iceberg; the memory and retrieval layers are what make AI actually usable in production.
  • For working professionals: if you own AI deployment at your company and notice AI keeps "repeating mistakes," the problem is almost certainly not the model — it's that the memory layer wasn't built right.
  • For consumers: next time you get an AI customer service call and it actually remembers what you said last time — don't be surprised, this is the frontier everyone's racing toward right now; but it also means your conversations are being recorded more structurally than ever.