Monday morning, and a client asks you to explain the brief again
At 10 p.m. last Wednesday, I was revising a client deliverable in my small Hangzhou studio. The next day’s online meeting opened with: “Didn't we already say not to exaggerate?” I said no problem, but I felt a pang of panic: the phrase “the best at understanding small businesses” on that PPT slide was one I had added myself.
If AI could remember requests like that, I wouldn't have to repeat them every time. My mistake was treating chat history as reliable memory. Tired people forget things, and so do assistants.
So-called “memory” is really three kinds of reusable information
Anthropic divides AI memory into three types: chat history, files, and runnable workflows. That framing feels practical. The point isn't to make AI remember everything forever; it's to put information that genuinely needs to persist in the right place.
I know Lin Yue, a friend who runs a corporate training account in Chengdu as a personal IP. Every day, she writes short-video scripts, replies to inquiries, and organizes client questions. She puts her clients’ “three biggest concerns” in a fixed document, records “avoid exaggeration, lead with the conclusion” in a handoff note, then lets the AI read them when needed. Her simplest habit is to click “Add file” at the start of every new conversation and select that client document.
That doesn't mean stuffing every chat into the system. My understanding is: chat history preserves context, files hold stable rules, and runnable workflows handle repeatable actions. For an ordinary founder, starting with the first two is usually enough.
What it cost to replicate: $0, about 30 minutes, no coding
Money: $0, using an existing AI chat tool. Time: about 30 minutes. Technical barrier: knowing how to copy, paste, and organize a document. First step: open the AI tool you already use, start a new conversation, and search for or enter a section called “Files,” “Projects,” or “Custom assistants.” If you can't find it, write your memory into a fixed document and paste it into each conversation.
I would start with only three short lists: client preferences and red lines, confirmed project decisions, and next steps. Keep each under one page and label it with a date; when records conflict, use the more recent confirmation. AI can still get an old detail wrong, so I wouldn't rely on it alone to “remember” prices, deadlines, or commitments made to clients. I still check those key points myself.
Not everyone needs this setup. If your work rarely carries across days, or every task starts from zero, it's completely fine not to build it.
Match the setup to your current workload
If I'm just starting out with scattered one-off requests, I'd begin with a one-page “personal preferences” sheet: tone, pricing boundaries, and my usual delivery format. The cost is low, and it's usually the easiest change to notice.
If I already have one or two stable clients, I'd keep a separate preferences-and-commitments record for each, then make “read this document before starting a new conversation” a fixed habit. That avoids carrying one client's habits into work for another.
If I'm scaling beyond a team of one, I'd add a layer of “shared team rules”: what everyone can see, what belongs only to a particular client, and what needs the lead's approval. Once the process is stable, I might let AI organize the records automatically, rather than building a complicated system on day one.