A marketing team can spend weeks teaching an AI how it works. Brand voice, customer groups, offers, approved claims, last quarter's decisions. Then someone opens a fresh chat and the briefing starts again from zero. The useful details are scattered across old conversations, meeting documents and one person's notes.
The real problem is the source of truth. Chat history belongs to a conversation, or to one person's account. Team knowledge needs a home people can open, correct and share. A small folder of plain-text files does that job because the contents stay readable without any special software.
So here is our position. Knowledge that lives only inside a chat is personal working memory. It has not become team knowledge yet.
Chat memory is useful, but it needs supervision
AI memory has improved. OpenAI says earlier saved memories could miss details that were never explicitly recorded, then go stale or wrong, and its June 2026 memory update was built to fix that. The admission is worth more than any promise that an assistant will simply remember everything.
Memory also varies by product and by setting. GitHub added cross-session memory for Copilot in JetBrains in August 2026, behind a settings toggle. Other tools handle continuity their own way. Your marketing setup should not depend on one vendor's recall rules.
Build a small folder that answers the recurring questions
Start with the files you would hand a capable new colleague on day one. Keep each one focused. Long manuals are hard to scan and easy to neglect. The folder can sit in a shared drive, a workspace or a repository, as long as the team controls access and your AI can read it.
- Start-here.txt. What the folder holds, which files matter for common tasks, who maintains them.
- Brand.txt. Positioning, voice rules, approved descriptions, words the brand avoids.
- Audiences.txt. Real customer groups, the questions they ask, and the evidence behind those descriptions.
- Offers.txt. Current offers, approved claims, exclusions, links to the official source material.
- Decisions.txt. Decisions with dates and reasons, including the ideas the team rejected.
- Meetings. Short notes with decisions and owners. Push old material into an archive folder.
Plain text does not mean unstructured. Use short headings inside each file, put the current guidance first, link out to the source documents. Tables belong in a spreadsheet if people need to sort them. The folder is an index of dependable context, not a copy of everything the team owns.
This structure is a suggestion, not a standard. Rename the files to match how your team searches. Keep stable facts apart from live campaign material, so an old launch date does not quietly walk into next month's brief.
Make every task start from the same source
A folder only works when the assistant actually gets it. Add a line to each reusable prompt naming the files to read and what to do when they disagree. Ask it to cite the file behind any brand, audience or offer claim. Missing evidence should produce a question for the owner, never a confident guess.
The instruction can be short. Before drafting, read Start-here, Brand, Audiences and Offers. Use current entries only. If a fact is missing, or two files disagree, stop and ask which source wins. That makes the knowledge check part of the job instead of a clean-up after the draft.
File-based instructions are already normal in technical AI work. GitHub Copilot code review can read a root AGENTS.md file automatically, and GitHub lets teams test instruction changes in a feature branch first. Marketing tools need their own setup. The operating idea transfers.
Keep the folder alive with small edits
Give every file an owner and a last-reviewed date. Update the file when a campaign changes an offer, when research changes an audience assumption, when leadership settles a disputed phrase. Do the edit while you are doing the work. A quarterly clean-up cannot recover decisions nobody wrote down.
Once a month, take two finished tasks and trace the important claims back to the folder. Fix what was missing, archive what expired, delete guidance that is now duplicated. Where two files argue, name the winner. You want a folder the team trusts enough to use, and stale certainty is worse than an honest blank.
Build the start-here file this week, then add only what one real task needs. If the wider knowledge pile is already a mess, a Second Brain Audit can sort what belongs in the shared source from what should stay elsewhere.