3.15 exposes AI poisoning — a business that went from Putian (莆田) to Silicon Valley

The problem: ephemeral brains

Consumer‑rights season (3.15) has once again shone a light on a quietly urgent technical problem: modern AI agents can't reliably remember. OpenClaw — the popular agent framework in China’s AI community — stores user memory in local Markdown files and repeatedly injects those notes back into the model. That works at first. Then tokens balloon, context windows fill, summaries lose narrative, and the agent repeats mistakes or drops task state. It has been reported that long‑running use can push MEMORY.md files into the tens or hundreds of entries, forcing truncation or ever‑rising token costs — and that means real products become brittle in the field.

The alleged fix: Engram and V4

Enter Liang Wenfeng (梁文锋) and the Engram idea. Liang’s paper and leaked V4 architecture reportedly propose embedding persistent memory into the transformer itself: a dedicated conditional memory space with O(1) hash lookup that does not consume the model’s context window or add per‑inference compute. If true, Engram would let models expand memory capacity nearly indefinitely while keeping latency constant. It has been reported that DeepSeek’s V4 will be an architecture‑level overhaul — claims range up to a 1‑trillion‑parameter, million‑context, native multimodal model with an April release — although those specifications remain unverified.

Why this matters — and the risks

Why should Western readers care? Because this is more than an engineering optimization: memory is now the difference between an assistant that helps and one that breaks workflows. Third‑party memory plugins such as MemGPT, Mem0, SYNAPSE and SimpleMem are useful stopgaps, but they leave the model “reading notes” rather than internalizing experience. Native memory could give Chinese firms a practical edge at a time when geopolitics — export controls on advanced chips and tighter U.S.‑China tech competition — make software differentiation crucial. But there are downsides too: persistent memories create new attack surfaces for data poisoning, privacy leaks and regulatory scrutiny. What happens if a model "remembers" manipulated or incorrect facts? Reportedly, that question is driving both intense engineering focus and growing concern among consumer advocates.

The broader takeaway: whether DeepSeek (reportedly with roots reaching from Putian (莆田) to Silicon Valley) succeeds or not, the community’s race from external plugins toward native memory will determine which agents are actually useful in production — and which will be exposed at the next 3.15.