How to give your AI a memory
Antonello da Messina, Saint Jerome in his Study, c. 1475
Vaughn DiMarco
Your AI forgets your business between sessions because nothing persists it. The fix depends on one question: who plugs into the memory: just you, a small team, or your customers through a product? Start free with built-in memory and a folder of plain text notes, then graduate to infrastructure like gbrain or Hindsight only when the work demands it.
Who plugs into the memory?
You have probably noticed the pattern. You open Claude or ChatGPT and spend the first ten minutes explaining your business again before you can ask for anything useful. Who your clients are, how you price, what you decided last month, why you decided it. The tool is smart and it knows nothing, because yesterday’s session is gone.
The AI world has a name for this: the persistent memory problem, and in 2026 it is where much of the interesting work is happening. Memory is what separates a tool you operate from a system that compounds. An assistant that remembers your conventions stops reinventing them every Tuesday. One that remembers your clients stops asking who they are.
The fix does not require anything exotic. The options have settled into a short list, and choosing between them comes down to a single question: who plugs into the memory?
Just you? A team of two to twenty? Or your customers, through a product you ship?
Each answer points at a different stack. Most bad memory setups come from answering this question wrong. A solo founder who buys team infrastructure pays for walls nobody needs, and a ten-person company sharing one brain with no walls leaks the first private note into everyone’s context.
If it’s just you
Start with what you already have. Claude and most serious AI tools now ship built-in memory: they keep notes on you between sessions and can search your old conversations. For a solo operator who works mostly in one tool, this is often enough, and it costs nothing. The real work is curation. Open the memory settings once a month, delete what is stale, correct what is wrong. An AI compounding on bad memories is worse than one with none.
The ceiling shows up when your work spans tools. What your chat assistant knows, your coding assistant does not, and neither can see the notes on your laptop.
The cheapest fix for that is a folder of plain text files. I recommend Obsidian for this: it is free, the files stay on your computer, they outlive any subscription, and every AI tool can read them. One file per client. One per project. One that says how you like things done. Point your AI tools at the folder and you have a shared reference layer for the price of a writing habit.
When you want more, the upgrade is called gbrain. It is an open-source “brain” built by Garry Tan, the CEO of Y Combinator, to run his own AI agents, and it sits on top of the same text files you already keep. It adds real search, a map of the people and companies in your notes, and overnight jobs that file new information while you sleep. The cost is commitment: it wants a computer that stays on, plus some setup you will mostly delegate to the AI itself. Run it for two weeks and judge it on one number: how often did asking the brain replace research you would have redone from scratch?
If it’s your team
One question decides everything here: does anything in the memory need to stay private?
If the answer is no, and you trust each other completely, the solo setup stretches. A shared folder of text files plus one brain on a small server works to about five people. Past that, or the moment the store contains client pricing, HR notes, or a candidate pipeline, you need permissions that the memory system itself enforces.
You have two ways to get them. Run it yourself with gbrain’s company mode, where each person gets their own slice of the brain and a query returns only what that person is allowed to see. Or buy it as a service from Hindsight, which was built for exactly this: many people, one brain, real walls, nothing for you to operate. If your team already lives in Notion, keep Notion as the place humans write. The memory layer reads from it and respects its permissions.
Whichever you pick, hold one rule: permissions live in the memory system and get checked on every query. Never put them in the AI’s instructions. An AI told to keep secrets is filtering by politeness, and politeness leaks.
If it’s your customers
A shorter section, because fewer of you ship AI products. If you do, memory stops being plumbing and becomes a feature. Each user gets their own isolated memory (Hindsight is the current production answer), and if the product should adapt to each person’s style and history, a layer called Honcho models people the way the others model facts. Your users will eventually ask two things: what does this remember about me, and can you delete it. Build both answers before launch, because retrofitting them is miserable.
What I would do Monday morning
For most owner-operators reading this, the move is smaller than the vendor landscape suggests.
1. Open the memory settings in your main AI tool. Delete what is stale or wrong. Twenty minutes.
2. Start the folder of plain text notes: clients, projects, one “how we work” file. Connect your AI tools to it.
3. Write decisions down when you make them, in one or two sentences, in those files.
4. Only after a month of that habit, consider infrastructure.
The order matters because tools churn and files last. Whatever memory system wins the next two years, it will read plain text, so every note you write today transfers. The founders who get compounding value from AI in 2027 are the ones whose businesses were writable in 2026.
We turned this decision into a two-minute interactive picker: answer four questions, watch the path light up, get the stack. Try it here. And if you want it applied to your specific business, tools and all, that is what the assessment is for.
Common questions
Why does my AI forget my business between sessions?
Because nothing persists it. Each session starts from an empty context window, so the model has no record of your clients, your pricing, or the decision you made last month. Memory is a layer you add, not a property the model has.
Do I need to buy memory infrastructure?
Usually not at first. Start with the memory already built into your main AI tool, plus a folder of plain text notes it can read. Only move to infrastructure like gbrain or Hindsight when the work outgrows that, which for most solo operators takes months.
How do I keep parts of a shared team memory private?
Put the permissions in the memory system and check them on every query. Never write the rule into the AI’s instructions instead. An AI told to keep a secret is filtering by politeness, and politeness leaks.
Find out what agentic workflows would save your team.
The $999 assessment identifies 5-10 hours per week of recoverable time, with a money-back guarantee if it doesn't.
Get the Assessment →