2.1 · Why Your AI Forgets
An AI model has no memory between conversations. Each new chat starts from zero. Inside one chat it only "remembers" because the whole conversation is sent back to it with every message.
That has two consequences.
It forgets you. Your name, your job, how you like things done. Unless something puts that back in front of it, it is gone.
You have felt this. You spend twenty minutes explaining your situation and get a great answer. The next day you open a new chat, ask a follow-up, and it answers as if it has never met you. So you explain it all again, a little shorter and a little more annoyed each time. After a few weeks you stop giving it the background at all, and the answers get generic. That is not you using it badly. It is how the technology works.
Long chats get worse, and cost more. As a conversation grows, the model has more to re-read each turn. Old details get diluted, and you use up your plan faster.
The usual fixes, and why they fall short
- Pasting a big "about me" prompt. Works once. You forget to paste it, and it never updates itself.
- Built-in chat memory. Handy, but it decides on its own what to keep, and it keeps a short list of loose facts rather than the full picture. You cannot easily see it, organise it, or correct it, and it is locked inside one app.
- One endless conversation. Slow, expensive, and eventually the early parts get summarised away.
- Custom instructions. Better, but it is one small box. It holds your tone and a few rules, not your projects, your people, or what happened last week.
All four share the same flaw: you are the one doing the remembering. The moment you get busy, the system stops working.
The fix that works
Put what the AI needs to know in plain files it reads at the start of every session, and let it update those files itself when it learns something.
That is all "memory" means in this vault:
- You can read it, because it is a normal note.
- You can fix it, because it is a normal note.
- It survives a new chat, a new model, and even a switch to a different AI, because it is only text in a folder you own.
The difference shows up fast. Tell a normal chat "draft the email to my landlord" and it asks who your landlord is, what the email is about, and how formal to be. Tell Jarvis the same thing and it already knows the landlord's name, that the lease renews in March, and that you like short emails. You gave it those facts once, weeks ago, and it wrote them down.
What belongs in memory
- Who you are and what you care about.
- How you want the assistant to behave, with the reason why.
- What is going on in your life and projects right now.
- Where things live: accounts, links, documents.
What does not belong
- Secrets. See 1.6 - Safety Rules Before You Give It Hands.
- Things that only matter today. Those go in the daily note.
- Things Jarvis can simply look up when it needs them.
Do this now
Think of one thing you have explained to an AI more than twice. Tell Jarvis, then say "remember that". You have just removed one repeat from your life.