Why Claude forgets your context every session (and the permanent fix)

By Alex Margolick, founder of Margolick AI · July 24, 2026 · 7 minute read

If you use Claude for real work, you already know the pattern. You spend twenty minutes catching the model up on your business, your pricing, your last three decisions. You get good output. You close the tab. Tomorrow you open a new chat and Claude knows nothing. Your pricing page and the supplier you dropped last week: gone. This is not a bug, and it is not going away. Claude is stateless by design. But there is a clean fix, and it does not require waiting for Anthropic to ship a memory feature.

Why Claude forgets: the technical reality

Every Claude conversation starts with an empty context window. The context window is the working memory the model uses during a single session. It holds your messages, Claude's responses, any documents you attach, and any instructions you give. When that session ends, the window is discarded. Nothing carries over to the next conversation.

This is not a limitation Anthropic is being slow to fix. It is a core architectural decision. Large language models process text in a window, and when the window closes, the text is gone. There is no database of "things Claude learned about your business" sitting on a server somewhere. Each conversation is independent.

The context window itself also has a size limit. Even within a single session, if you paste in your entire business history, product catalog, and every past conversation, you will hit that ceiling. And as the window fills up, the model's attention on any single piece of information gets thinner. More context does not always mean better answers.

The workarounds people try (and where each one breaks)

Claude Projects

Claude's Projects feature lets you attach documents to a workspace so every new conversation inside that project starts with those documents loaded. This is genuinely useful. It means you can write up your business context once and not paste it every time.

But Projects are a delivery mechanism, not a memory system. The documents inside a project are static until you manually update them. If your pricing changed last Tuesday and you forgot to update the project doc, Claude will confidently work from the old numbers. Projects also do not capture what happened in previous sessions. They hold reference material, not session history or decisions made along the way.

Custom instructions and system prompts

Both Claude and ChatGPT offer a place to write standing instructions ("I run a DTC skincare brand, always respond in a direct tone, prices are in USD"). This helps with tone and basic identity, but the character limits are tight. You cannot fit a real business context into a few hundred characters. And like Projects, these instructions are static. They go stale the moment your business moves.

Manual pasting

The brute force approach: keep a document somewhere and paste it into the first message of every new chat. This works with any model, which is its one real advantage. The downsides are obvious. It is tedious, it burns context window space on backstory instead of the actual task, and it still goes stale unless you discipline yourself to update the document after every session.

ChatGPT's built-in memory

ChatGPT has a memory feature that saves fragments across conversations. It remembers things like "user prefers bullet points" or "user's company is called Acme." This sounds like the fix, but it has three problems. First, ChatGPT decides what to remember, not you. You cannot audit the full list of what it "knows" about your business or correct something it inferred wrong. Second, the memories are shallow fragments, not structured business context. "User sells candles" is not the same as knowing your current inventory levels, your Q3 pricing, and the decision you made last week to drop wholesale. Third, none of it transfers. If you switch to Claude or Gemini for a task, that memory stays locked inside ChatGPT.

Gemini's approach

Google's Gemini can read your Google Drive, which gives it access to whatever you have stored there. This is powerful if your business already lives in Google Docs, but it creates a different problem: Gemini can see everything, with no structure telling it what matters right now versus what is outdated. A six-month-old strategy doc and today's pricing sheet look the same to the model unless you tell it otherwise.

The real fix: a structured context layer

Every workaround above fails for the same reason. They treat context as something you push into a chat. The fix is to treat context as something that lives outside every chat, stays current on its own, and gets read at the start of every session by whatever model you are using.

I call this a context layer. It is a small set of pages, stored in Notion or Google Docs or a plain text file, that answer one question: "What does my AI need to know to pick up where the last session left off?" The pages are yours, not locked inside any vendor. Any model that can read text can use them.

What goes in the context layer

Four things, and nothing else:

1. A current-state page. Where the business stands right now. What you sell, current prices, active projects and their status, open blockers, next three actions. Nothing historical. If a fact on this page is no longer true, it gets corrected immediately. Keep it under a thousand words.
2. A decisions log. Dated entries, one line each. "2026-07-10: Dropped the $29 tier, margin too thin after fees." Without this, a fresh Claude session will cheerfully reopen debates you settled weeks ago, and argue the losing side convincingly.
3. Session close notes. At the end of every working session, the AI writes down what was done, what changed, and what the next session should start with. This takes two minutes. Skip it and the current-state page goes stale within days.
4. Permission rules. What the AI may do on its own and what always waits for you. "Draft anything, but nothing goes to a customer without my approval. Never change prices or sign agreements." Written rules matter more as models get more capable and you give them more autonomy.

How the open and close ritual works

Every session starts with a read. The AI opens the current-state page and the decisions log, confirms what it understood, and begins working from today's facts. No recap needed. No "let me catch you up." The context is already there.

Every session ends with a write. The AI drafts close notes, updates anything on the current-state page that moved, and logs any decisions made. Tomorrow's session starts by reading what today's session wrote. The loop is self-maintaining.

This works because the context lives outside any single model or vendor. If you use Claude for strategy on Monday and ChatGPT for copywriting on Tuesday, both read the same pages. If you try Gemini next month, it reads the same pages. The memory belongs to your business, not to any chat window.

What this changes in practice

The obvious win is continuity. Your AI picks up where the last session stopped, every time, without you re-explaining anything.

The less obvious win is cost and speed. Re-pasting business history into every session is expensive. My session starts used to carry tens of thousands of tokens of accumulated backstory. With a current-state page, that dropped to about a thousand tokens. Sessions start faster, responses stay sharper because the model is working from a clean snapshot instead of wading through stale history, and on API pricing the difference shows up on the bill.

There is also a resilience benefit. If you get interrupted, lose your train of thought, or just close the laptop, the state survives. Any new session rebuilds from the pages. I built this system because my own brain drops threads constantly, and the same property makes it work for a team: two people (or two models) can work the same business context without a handoff meeting.

Frequently asked questions