The gap between an AI writing tool and a writing agent is memory, and most builds skip it

Spent a while building writing help into a product, and the thing I underestimated: a writing tool and a writing agent are not the same thing, and the difference is almost entirely memory. A writing tool is stateless. You give it text, it gives you better text, done. Genuinely useful, and most "AI writing" features are exactly this. But every request starts from zero. It does not know how you write, what you have already published, or that you corrected the same thing last week. So it makes the same mistakes on repeat and you re-teach it every session. A writing agent is the same generation ability plus persistence. It remembers your style decisions, the terms you always change, the topics you have already covered so it stops repeating them, and the feedback you gave on past drafts. That memory is what turns it from a fancy autocomplete into something that gets more useful the longer you use it. What I learned building the memory layer: - Store decisions, not full history. "User always changes X to Y", "avoids the word Z", "target reader is A" beats dumping every past draft into context. Cheaper, and it actually steers output. - Separate durable style memory from per-project memory, or a one-off tone for one client bleeds into everything. - Let the user see and edit what it remembers. Silent memory that is subtly wrong is worse than no memory, because you cannot tell why the output drifted. Most builds I see stop at the stateless tool because it demos well. The memory is unglamorous, and it is the whole difference in daily use. Anyone gone deep on the memory side of a writing agent? Curious what you store and what you deliberately throw away.