Why AI forgets what you told it

People describe the same experience in the same words: it forgets me. The details do not carry over. Every morning I have to re-explain the project, the client, and what we decided last week, as if the tool and I had never met.

The mechanics are unglamorous. A chat model reads a context window, which is a fixed budget of text it can see at one time. Your conversation, any files you attach, and any memory the product injects all compete for that budget. When the conversation gets long, the earliest parts fall out of view. When you close the session, the whole window is discarded. And at no point did any of it change the model: the model you talk to tomorrow is the same one everyone else talks to, with no trace of your Tuesday in it.

So the forgetting is not a malfunction, and it is not personal. The product was built to hold a conversation, not to hold your work. Anything that needs to survive the session has to live somewhere outside it.

Partly. ChatGPT and Claude both ship memory features now, and they are genuinely useful for what they hold: your preferences, standing instructions, a compressed sketch of past conversations. If the problem is re-typing the same instructions, memory features help.

Two limits show up in practice. The first is depth. What gets kept is a summary the product chose, not the decisions, sources, and reasoning your work actually runs on, and you cannot fully see or edit what it kept. The second is where it lives. That memory sits inside one vendor’s login. It does not follow you to another tool, and when you cancel the account, it does not come with you. People who hit this boundary describe it plainly: you don’t really own it.

The observation this page rests on: memory inside the same login is still their memory, not yours. It makes one product more convenient. It does not give your work a memory.

Keep the memory in files you own. Plain text, on your machine, readable by any model. Instead of hoping a product retains the right things, you write down the things that matter once, and hand them to whatever tool you are using at the moment of work.

Three folder shapes cover most of what a tool needs to stop meeting you cold:

• Decisions. What you already chose and why. Pricing calls, tool choices, the client boundaries you set in March. This is the file that stops the tool from re-litigating settled questions.
• Exceptions. Where the rule bends. The one client on old pricing, the process step you skip for rush orders. Exceptions are what a generic assistant can never guess, and they are usually where its advice goes wrong.
• How we work. The way things actually get done: who approves what, what a finished deliverable looks like, what you refuse to do. One page is enough to start.

Each file is short, written once, and updated when reality changes. Load them at the start of a session by pasting or attaching, or connect the folder so the model reads it directly. Kept and grown deliberately, that set of files becomes owned, portable working memory: what you know about your work, in plain files, readable by any model. A personal AI brain is this idea carried to its full shape.

This is where the file approach stops being a preference and starts being a strategy. Vendor memory does not transfer. Move from ChatGPT to Claude, or to whatever ships next quarter, and you start from zero: a cold start with a tool that has never heard of your business, your clients, or your decisions.

Files do not have that problem. The same decisions, exceptions, and how-we-work files brief any model equally well, because every model reads plain text. Switching tools becomes a choice about capability and price, not a choice about abandoning your history. The better the models get, the more that folder is worth, because each new model starts from everything you wrote down instead of from nothing.

The same logic scales from one person to a company. If a team’s accumulated context lives inside one platform’s accounts, it is trapped there. How to keep the company’s context in company hands is its own question, covered in owning your AI context.