> ## Content Index > Fetch the complete content index at: https://www.appliedcomms.ai/llms.txt > Use this file to discover other available public pages before exploring further. # How to Give AI a Memory: A Practical Guide for Comms Teams - URL: https://www.appliedcomms.ai/give-your-ai-a-memory/ - Published: 2026-09-08T06:57:28.000Z - Updated: 2026-09-08T06:57:28.000Z - Description: A lot of AI output feels generic, and one reason is that the model has no idea what your organisation has already said, approved or decided. Here is where a durable memory can live, ten practices that keep it trustworthy, and three things you can do today. - Author: Michael MacLennan - Tags: Guides, Experiments & Applied Learning ## **The Brief** - **The problem:** A lot of AI output reads as generic, and a primary reason is that the model has no idea what your organisation has already said, approved or decided. It writes like it has just walked in the door... because it has. - **The fix:** A maintained memory the AI reads at the start of every session and that someone updates as decisions land. There are four places it can live: the lightest of them is free and takes an afternoon. - **The takeaway:** 10 practices below, grouped from light to heavy, plus three things you can do today. Casey Newton has written up a month working with the ambitious version; I have been running the comms version for the best part of a year. --- ## **The Story** Back in April, the AI researcher Andrej Karpathy [posted an idea](https://x.com/karpathy/status/2039805659525644595): use an LLM to build and maintain a personal knowledge base, a folder of markdown files the model organises and updates as you feed it new material. Casey Newton built one, and his [account of living with it](https://www.platformer.news/karpathy-llm-wiki-journalism-productivity/) is the most concrete I have read. His wiki now runs past 1,400 pages, seeded with six years of Platformer archives, updated every morning, generating timelines and story ideas, and demanding, in his words, the maintenance of a vintage sports car. Newton is a journalist, so his wiki tracks a beat. The problem it solves is one every comms team knows: institutional memory. What have we said publicly on this before? Which phrasing did legal approve? What did we decide in March, and why? In most organisations that knowledge lives in people's heads, an unsearchable shared drive and a departed colleague's inbox. **When teams explain to me their AI produces generic output, missing context is usually a large part of the complaint.** If this sounds at all familiar, the encouraging part is that fixing it needs no technical skill, just the editorial judgement comms people already have. However, it does need a decision about where the memory sits, which is where most teams get stuck. \[A quick definition first, because the word gets stretched. By memory I mean a maintained set of instructions, facts and decisions your AI can refer to across tasks. Attaching a document is not the same thing: you still need a way to supply that context each time, check it was used and approve changes to it, and the tools below handle those steps differently.\] ### **Where the memory actually lives** The first three options are places to keep maintained context, in rough order of effort. The fourth is different, because it writes itself. They are not exclusive, and most people end up running two. **1\. Inside the chat tool, as a project workspace.** [ChatGPT Projects](https://help.openai.com/en/articles/10169521-projects-in-chatgpt), [Claude Projects](https://support.claude.com/en/articles/9517075-what-are-projects) and [Gemini Gems](https://support.google.com/gemini/answer/15146780) all give you a set of standing instructions plus attached files, all work on free accounts with limits that rise with the plan, and ChatGPT projects can now be shared with colleagues. If you are starting this week, start here. Nothing to install, and you can test the idea properly in an afternoon. **2\. On your own device, as a folder of files.** Plain markdown, one file per subject, read by an assistant at the start of a session and written back to at the end. This is Karpathy's original shape and it is the one I run for the [Faur](https://faur.site/) / [Comms With AI](https://www.commswith.ai/) ecosystem. Everything stays in text you own, it moves between tools, and you can see exactly what changed and when. The catch is that you need a desktop assistant allowed to read and write local files, a bigger step than most comms teams have taken. Also, it can be a frustration when you're out and about without access to your hard drive, and an urgent task comes in. (For this reason, I'll sometimes create a 'light' cloud Projects Folder in another GPT tool, meaning that I can do a decent standard of work away from home on this, regularly updating its memory.) **3\. In the systems the organisation already runs.** [SharePoint agents](https://support.microsoft.com/en-us/office/get-started-with-agents-in-sharepoint-69e2faf9-2c1e-4baa-8305-23e625021bcf) grounded on a site or document library, [Copilot Notebooks](https://support.microsoft.com/en-us/microsoft-365-copilot/how-microsoft-365-copilot-notebooks-works) scoped to a curated set of references, or [company knowledge in ChatGPT](https://help.openai.com/en/articles/12628342-company-knowledge-in-chatgpt-business-enterprise-and-edu) on Business, Enterprise and Edu plans. This is where a whole team's memory eventually belongs, and it raises the stakes. These agents inherit the permissions of the library behind them, so anything the library was quietly over-sharing gets surfaced at organisational scale, quickly and cheerfully, by a chatbot. **4\. The memory that builds itself.** [Claude's memory](https://support.claude.com/en/articles/11817273-use-claude-s-chat-search-and-memory-to-build-on-previous-context) keeps a separate summary per project, lets you edit what it has stored, and stays off for Team and Enterprise until an owner enables it. ChatGPT has its equivalent. I leave mine on and find it useful, though as a convenience alongside the real thing: the model writes it, you rarely read it, and you cannot hand it to a new starter on their first morning. ### **What mine looks like** A glimpse behind the curtain at option two, then. Across [Faur](https://faur.site) and [Comms With AI](https://commswith.ai) I keep a set of operating notes rather than a wiki: plain markdown files, one per part of the business, holding the current position, the conventions and the decisions with the reasoning behind them. Each client and product has its own working-memory file, and Claude reads the relevant ones at the start of a session and updates them as decisions land. The practical effect is that I stopped re-explaining my business every morning. ![](https://storage.ghost.io/c/e9/fc/e9fc5e96-e3eb-42e1-81da-72a4c05345f0/content/images/2026/09/Screenshot-2026-09-08-at-07.40.21.png) Here's the beginning of the H2 2026 Comms With AI strategy document, which has been updated at semi-regular intervals to align with latest events as well as with the wider suite of foundational documents for the wider Faur ecosystem The example that matters most is approved language. For one client, a large-scale infrastructure project, the strapline, the approved short description and a set of key facts and figures all had to be exactly right, because stakeholder approval depended on them. Those live in a playbook, and the standing instructions tell the assistant to go to that playbook whenever it produces anything for that client. The instructions say where to look; the documents hold the wording. Separately, a scheduled sweep runs over recent output looking for slips in the approved wording, tone of voice and other agreed conventions. Keeping the language available is the starting point; checking it is being used is part of maintaining the memory. Limits to be aware of (I appreciate that Newton is also honest about his!). The files need gardening: they grow long and have to be condensed, and two notes will drift out of agreement unless something reconciles them. The model's own writing habits creep in unless you police them, which is why Newton had much of his system rewritten into AP style and why my notes carry a standing instruction about em dashes and other AI giveaways that still gets tested weekly. And a remembered claim is not a verified one. Newton says he opens the original sources to check his wiki for hallucinations, and that habit matters more than anything else in his piece. ### **Ten things that make it work** These are the practices that separate a memory people trust from a folder nobody opens. The first three are worth doing whatever your setup. ![](https://storage.ghost.io/c/e9/fc/e9fc5e96-e3eb-42e1-81da-72a4c05345f0/content/images/2026/09/Screenshot-2026-09-08-at-07.48.04.png) My root instructions for Faur contain detail on facts and figures as well as more mundane conventions for keeping filing in decent order, which can help massively with the day to day **Light: start here** 1. **Write instructions before you build a library.** One page of "how we do things here" does most of the work, and in my experience models follow standing instructions far more reliably than they retrieve from attached documents. Then add the reference material a task actually needs, because instructions and evidence do different jobs. 2. **Split the durable from the dated.** Two files rather than one: how we work, which is stable for months, and what is live now, which changes weekly. Mixing them is how an AI ends up confidently quoting last quarter's position at this quarter's journalist. 3. **Turn corrections that should apply again into a line in the file, the same day.** When you fix the output, fix the memory, but only for the fixes you want repeated. One-off preferences pile up into contradictory rules. My em dash line exists because I got tired of removing them by hand. **Regular use: the structure that keeps it trustworthy** 1. **Decide what never goes in it, and check where it goes.** A memory file is a disclosure surface. Embargoed announcements, personal data, live legal advice, an unannounced restructure: write one line at the top of every file naming what stays out, and brief anyone who can edit it. A never-list is a reminder rather than a control, so the real protection is working in a tool your organisation has approved, with material authorised for that environment. A file on your laptop is still processed wherever the model runs. The [governance templates](https://www.commswith.ai/library/governance/) in the Comms With AI library include an AI use policy you can adapt rather than draft from scratch. 2. **Name one owner and put the 'gardening' in the diary.** Thirty minutes a fortnight, one named person, and a note at the bottom of each file recording when it was last amended. The easy way to run the session is to ask the assistant what in the notes is out of date and accept or reject its suggestions: machine drafts, human approves. Newton's vintage sports car line is the right warning. An unmaintained memory is worse than none, because people carry on trusting it. 3. **Give each client, brand or product its own memory.** Separate files, or separate projects, keep each context sharp and keep confidential work where it belongs. One memory covering the lot will answer everything vaguely. 4. **Record where approved language came from.** For every current position or approved phrase, note the source or approval reference, who approved it, when it took effect and when it needs review, and whether it is proposed, approved or superseded. When two notes disagree, approval beats recency: a fresh draft never silently outranks an older approved position, and superseded wording is marked rather than deleted, so the reasoning survives. **Heavy use, and teams** 1. **Record what you rejected, and why.** The rarest file in any team's memory and the most valuable. "We considered this in March and did not do it, because." It stops you relitigating settled arguments and it stops your AI enthusiastically proposing the thing legal killed six months ago. 2. **Run a cold-brief test once a month.** Hand the assistant a real task using only the memory and no verbal explanation, then judge the output against what a competent new starter would produce. If it fails, treat it as a diagnosis rather than a verdict: the gap could be missing context, a document it did not retrieve, an unclear instruction or a limit of the model, and each needs a different fix. 3. **Move from files to sources only when files stop scaling.** Connecting an agent to SharePoint or Drive is the right answer eventually, and it changes the risk. The agent sees whatever the underlying permissions allow, so run the permissions check before you point anything at the library, not after the first awkward answer. [Last Thursday's piece on the Hugging Face incident](https://www.appliedcomms.ai/lessons-from-openais-hugging-face-break-in-why-orchestration-is-your-next-must-have-comms-ai-skill/) argued that orchestration, briefing agents well and writing down when they should stop and ask, is a comms skill. Memory is the other half of that argument, since a brief is only ever as strong as the context sitting behind it. --- ## **The Practice: what you can put into practice today** 1. **Choose one task you are tired of re-explaining.** A familiar media enquiry, a stakeholder update, a standard boilerplate paragraph. One task, not a taxonomy of your whole team's knowledge. 2. **Build a small context pack for it.** In a tool your organisation approves, open a project and give it one page of working instructions and one current, approved reference, with the source, the approval date and the owner written on it. Add a line at the top naming what must never go in. Use only material authorised for that environment. 3. **Test what changes.** Give the same task to a blank chat and to the project, then compare factual accuracy, use of approved language and how much editing each needs. It is a first signal rather than proof, and any mistake is worth investigating before you add more material. What is the one thing you are tired of explaining to your AI? Take a moment to reflect, ruminate, then put the solution in place – it's important to maintain your own memory of how and where you're making progress with AI implementation. *Every Applied piece follows the same shape: The Brief, The Story, The Practice.* **More like this** - [How Project Folders Supercharged My AI Comms Workflow](https://www.appliedcomms.ai/project-folders-ai-comms-workflow/): the 2025 piece on ChatGPT and Claude projects as a working memory, which this guide builds on. - [The living fact layer and the death of the PDF: NOAN CEO Neal Mann on why most AI-powered communications is built on sand](https://www.appliedcomms.ai/noan-neal-mann-interview/): the interview that made the case for maintained facts over attached documents. - [From Content Creator to System Builder: Claude Cowork Just Changed What Comms Professionals Can Actually Do With AI](https://www.appliedcomms.ai/claude-cowork-review/): the folder-of-files setup described in option two, from the week it started. --- ![](https://storage.ghost.io/c/e9/fc/e9fc5e96-e3eb-42e1-81da-72a4c05345f0/content/images/2026/09/image.png) ## **Coming up: Leader Comms With AI Webinar With Stephen Waddington** I'll be putting some of these questions to Stephen Waddington live on Wednesday 16 September, in the next Comms With AI Leader Interview: [Outputs or Outcomes? Stephen Waddington on the comms teams getting AI right](https://www.eventbrite.co.uk/e/outputs-or-outcomes-stephen-waddington-on-the-comms-teams-getting-ai-right-tickets-1993691162937). Registration is free.