Fudan NLP releases A5, an assistant that remembers your style and takes initiative on its own

Researchers at Fudan NLP, the natural language processing group at Fudan University in Shanghai, have unveiled A5, a proactive assistant that builds a memory of the user and offers help on its own. The project addresses two complaints familiar to anyone who works with chatbots regularly: the assistant forgets everything when you open a new dialogue, and it never makes the first move.

The cream that melts in a new chat

The first problem will be familiar to anyone who has written with the help of an LLM. Model-generated text is often full of clichés and defensive hedging: every claim is wrapped in "on the one hand" and "it should be taken into account." To get natural, human-sounding language, you have to spend a long time tuning the prompt, explaining what exactly irritates you, and supplying examples.

Then you open a new dialogue, and everything starts over. The authors of A5 compare this to cream that spreads out: the style you built does not hold its shape. The familiar "AI flavor" returns, and with it comes the manual work of scrubbing it out again.

A5 targets exactly this gap. The assistant is supposed to accumulate knowledge about how the user writes, what they accept and what they reject, and carry that knowledge from session to session. The longer you work with it, the more accurately it should hit your tone on the first try.

Help that arrives before the question

The second problem is subtler. The authors describe it with an everyday example. You tell the assistant your paper has been accepted at a conference in Europe and ask it to gather materials for the trip. An ordinary chatbot congratulates you, produces a list of documents and preparations, and the story ends there. After that, everything depends on you: remembering, opening the chat, and asking again.

A5 is built around the idea that a useful helper keeps context in mind for longer than a single reply. A month before departure, it comes back to the topic by itself. It can remind you about the visa, the booking, or registration deadlines, because it knows the date and purpose of the trip. No prompt and no repeated explanation needed.

This is a change of role. A classic chatbot works like an information desk: it answers a question and forgets it. A proactive assistant is closer to a secretary who manages your affairs over time.

What is claimed for A5

• Developer: Fudan NLP, Fudan University, Shanghai.
• Type: a personal assistant with long-term memory.
• Memory: language and style preferences persist between dialogues.
• Behavior: initiates help on its own, based on accumulated context.
• Main scenario: preparing for an event where deadlines matter, for example a business trip to a conference.

Where proactivity can break down

The idea is clear, but it comes at a cost. Memory about a user means storing personal data: plans, habits, work details. The better the assistant knows a person, the sharper the questions become about where that data is kept, who has access to it, and how to delete it.

The second risk is intrusiveness. An assistant that reminds you at the wrong time or about the wrong thing quickly becomes annoying and gets switched off. The line between "thoughtfully warned" and "spamming notifications" depends on the quality of the model of the user's priorities. Errors in memory are dangerous too: if the assistant locks in a chance remark as a permanent preference, it will replicate that remark in every text.

For the industry, A5 is interesting as one more signal: competition is shifting from the quality of a single answer to the quality of the relationship with the user over time. Memory and initiative are becoming product features rather than side effects of a long context window. Meanwhile, academic groups like Fudan NLP can afford to explore these mechanics openly, while commercial players stay cautious because of privacy risks.

How to evaluate such systems in practice

If you are evaluating an assistant like this for yourself or your team, check a few things.

• Can you see and edit what the system has remembered about you?
• Can you quickly erase the memory entirely or by topic?
• Is the frequency of proactive messages configurable?
• How well does the style hold up on new types of tasks, not only the ones where you trained it?

If the answers are yes, such an assistant can genuinely save time on repetitive setup and small-scale planning. If not, you are looking at yet another chatbot that remembers more than is comfortable for you.