Every chat starts from a blank page
The assistant only sees what is in the current conversation window. Close the chat, open a new one, and the model knows nothing about the past. It did not forget: that information simply is not on its input. Hence the feeling that you hire a new colleague every morning and brief them on the project from scratch.
Built-in memory answers a different question
ChatGPT and Claude do have their own memory, and it genuinely helps. But it has two limits you hit the moment work involves more than one person. It belongs to one account: a colleague cannot see it and will still start from zero. And it belongs to one model: move to another assistant and the memory does not come with you.
A rules file remembers how, not why
An instructions file for the agent is useful and nobody is arguing against it: it sets the rules of work. But it is static, written by hand, and above all it answers a different question. Rules say how to do things. They do not hold what was decided six months ago and why the previous option was dropped. That reason is the first thing to disappear.
What actually closes the gap
A memory layer outside the assistant that the agent reaches into by itself. Three properties make it work instead of becoming one more place you have to remember to check.
- Stored separately from the provider, so it survives a change of assistant and is available to more than just you.
- Written by the agent, not by you: the decision is saved as it is discussed, together with the reason.
- Returns a list first and reads in full only what is needed, otherwise memory eats the context window faster than it pays for itself.
How to test this on yourself
Take a decision you made a month ago and ask a fresh chat about it. Not what you did, but why you picked that option. If there is no answer, you reconstruct it from your own head every time. That is the work you should not be doing.