How to stop Claude and ChatGPT from forgetting your business between sessions
Claude and ChatGPT forget your business between sessions because every new chat starts with an empty context window. The fix is external memory: a small set of written pages that live outside the chat and get read at the start of every session. Four pages are enough. A current-state page that says where the business stands right now. A decisions log that records what you decided and why. Session close notes, written in the last two minutes of every chat. And plain permission rules that say what your AI may do without asking you first. Keep those four current and any model you open picks up exactly where the last one stopped.
This guide walks through the whole setup. Building it from scratch takes about an hour in Notion, Google Docs, or a plain text file. Maintaining it takes about two minutes per session. I run my own company this way, so everything below comes from daily use rather than theory.
Why does your AI forget everything between sessions?
Large language models are stateless. A chat session holds everything you have said inside a context window, and when the session ends, that window is gone. Open a new chat and the model knows nothing about your pricing change last Tuesday, the supplier you dropped, or the plan you agreed on at midnight. It is not broken. It was built that way.
Both major vendors have shipped partial fixes. ChatGPT has a built-in memory feature that saves fragments across chats, and Claude has Projects, which keep documents attached to a workspace. Both help. Neither is enough on its own. ChatGPT's memory decides for itself what to keep, so you cannot audit what it knows or fix what it got wrong. Claude Projects hold documents but the documents still have to exist, stay current, and say the right things. And nothing transfers between vendors: the context you build up inside ChatGPT is invisible to Claude and Gemini.
So the working answer is the same one businesses used long before AI: write the important things down in one agreed place, and make reading that place the first step of every working session.
What is an external memory system?
An external memory system is a small set of documents, owned by you and stored outside any chat, that holds the current state of your business in a form an AI can read in seconds. It replaces the twenty minutes of "let me catch you up" that otherwise opens every session. The AI reads the pages, confirms what it understood, and starts working from today's facts instead of a blank slate or a half-remembered summary.
The version below has four parts. Each one exists to stop a specific failure I kept hitting while running my business with AI as the main operator.
The four-part setup
1. Write a current-state page
One page, always current, that answers the question "where does this business stand today?" Mine covers what the company sells, current prices, active projects with their status, open blockers, and the next three actions. Nothing historical, nothing aspirational. If a fact on the page is no longer true, it gets corrected or deleted the moment you notice.
The discipline that makes it work: this page is a snapshot, never an archive. The moment it grows past what an AI can read in one gulp, it has failed. Mine stays under a thousand words. When a session starts, the AI reads this page first, and that single read replaces the entire recap.
2. Keep a decisions log
A dated list of decisions, each with one line of reasoning. "2026-07-14: Dropped the $29 tier. Margin too thin after fees." That is a complete entry.
The log solves the most expensive failure mode of working with AI over weeks: relitigating settled questions. Without a log, a fresh session will cheerfully reopen a pricing debate you closed a month ago, and it will argue the losing side well. With a log, you point at the entry and move on. It also stops the quieter failure where the AI follows an old plan you have since changed, because the change is written down where the next session will see it.
3. Close every session with notes
The last two minutes of any working session belong to the close. Ask the AI to write down what was done, what changed, what is unfinished, and what the next session should do first. Then have it update the current-state page if anything on it moved. That is the whole ritual.
Skip this step and the system dies quietly. The current-state page drifts from reality, sessions start trusting a stale snapshot, and within a week you are back to re-explaining. Close notes are cheap insurance: two minutes at the end of each session keeps every future session's first minute honest.
4. Write permission rules for your AI
A short page that says what the AI may do on its own and what always waits for you. Mine boils down to a few plain sentences: draft anything, but nothing goes out to a customer without my approval. Never change prices, sign anything, or spend money. When unsure whether something is allowed, ask.
The single most useful rule I run: the AI never edits the canonical pages directly. It stages proposed changes for review, and I promote them. One sentence of policy, and it removes the entire class of "the AI quietly rewrote my source of truth" failures.
Written permission rules matter more as models get more capable. An assistant that can act needs the same thing a new employee needs: a clear description of the job's boundaries, in writing, that it reads before starting.
How do you get your AI to actually read it?
Three ways, in order of convenience:
- Direct connection. Claude and ChatGPT can both connect to Notion and read pages themselves. Tell the AI where the pages live and make "read the current-state page" the standing first instruction of every session.
- Project attachment. Put the pages in a Claude Project or reference them in ChatGPT custom instructions, so every new chat inside that workspace starts with them loaded.
- Paste. The zero-setup option. Copy the current-state page into the first message. Crude, and it works with any model ever made.
Whichever route you pick, the open ritual mirrors the close: the session starts by reading state, and ends by writing it.
What does this change in practice?
The obvious win is continuity. The less obvious win is cost and speed. Re-pasting history into every session is expensive: my session starts used to carry tens of thousands of tokens of accumulated backstory. With one current-state page read at open, that dropped to about a thousand tokens, roughly ninety percent less context per session. Sessions start faster, responses stay sharper because the model is not wading through stale history, and on metered API pricing the difference shows up on the bill.
There is also a resilience win that matters if your attention is the bottleneck. If you get interrupted mid-session, or your focus drops the thread at 2pm, the thread survives, because it was never only in your head. Any new session, tomorrow or in ten minutes, rebuilds from the pages. I built this system because my own brain works that way, and the same property is what makes it work for a team: the memory belongs to the business, not to any one chat, model, or person.
Common questions
Does ChatGPT's built-in memory solve this?
Only partly. Built-in memory stores scattered fragments it chooses on its own, and it does not transfer to Claude or Gemini. An external system is complete, readable, under your control, and works with every model.
Do I need Notion for this?
No. The method works in Google Docs, Obsidian, or a plain text file. Notion is convenient because Claude and ChatGPT can connect to it and read the pages themselves.
How long does the upkeep take?
About two minutes per session, spent on close notes and any change to the current-state page. The setup itself takes around an hour from scratch, or about ten minutes with a pre-built template.
What happens if I skip the session close notes?
The current-state page goes stale within a few days and your AI starts working from old facts. Close notes are the step that keeps the whole system honest.
Will this work with Gemini or a local model?
Yes. The pages are plain text, so anything that can read text can use them. That includes Gemini, local models, and whatever ships next year.
If you would rather copy than build
Everything above is buildable by hand, and the free guide covers the one rule I would add first: the fire-no-write axiom, which keeps an AI from quietly breaking your source of truth.
If you want the whole thing pre-built, the AI Cofounder OS is this exact system as a ready-made Notion setup: cockpit, decisions log, open and close checklists, permission rules, and a worked example business. It starts at $39 with a 30-day refund, and there is a Counterweight edition built for founders with an ADHD-like brain.