We note that MemPO, open-sourced by a Chinese team this week, deserves attention from traditional industries: it is a training framework that teaches AI Agents (AI capable of autonomously completing multi-step tasks) to write their own "work notes," aiming to solve the engineering challenge of long workflow tasks constantly "losing the thread."
What This Is
MemPO's core idea is to turn "memory" into a trainable variable, using a dual-channel reward mechanism (scoring from two dimensions simultaneously): one reward channel judges whether the answer is correct (1 for correct, 0 for wrong); the other rewards "note quality"—letting another model answer questions using only your notes, giving reward if it still gets them right and penalizing if it cannot.
This design lets the model learn on its own "what is worth recording and how to record it," with no need for developers to manually design memory modules. The code is open-sourced on GitHub and Hugging Face.
Industry View
Optimistic voices argue this hits the real bottleneck for Agent deployment—not that models aren't big enough, but that long-workflow stability is lacking. Equally worth our attention are the negative views: first, research code typically needs 1–2 more years to become enterprise-grade tools; second, the reward mechanism carries a risk of being "gamed," where the model could learn to write notes that appear useful but contain no substantive content to trick the reward; third, some researchers argue that expanding single-pass context to the million-token level could bypass the memory problem without dedicated optimization.
Impact on Regular People
For enterprise IT: when evaluating whether AI Agents can handle real business workflows (such as customer service tickets, document review), prioritize testing long tasks of 5+ steps—good performance on short tasks does not indicate production readiness.
For individual professionals: today's AI assistants are still only competent at short tasks; letting them independently complete multi-step work like "research—organize—write report" is not yet realistic.
For the consumer market: consumer-facing Agent products (auto-planning trips, auto-organizing email) will only become truly usable in 2–3 years; there's no need to pay for the "fully autonomous assistant" concept today.