Why your AI forgets, and what "memory" really means.
Context windows, chat memory and a company's records are three very different things. Knowing which is which explains most of the forgetting.
"It forgot what I told it" might be the most common complaint about AI. You set something up carefully at the start of a conversation, and an hour later it's acting like it never heard it. Or you start a new chat and have to explain everything again.
The word "memory" gets used for three completely different things, and once you separate them, the forgetting stops being mysterious.
1. The context window: short-term attention
The context window is everything the model can see right now: your instructions, the conversation, any documents, the results of whatever it just did. It's big, but it isn't infinite, and the model doesn't pay equal attention to all of it.
When a conversation gets long, one of two things happens. Either the oldest parts fall off the edge, or the system summarises them to make room. Both lose detail. That instruction from the start? It might be summarised into one vague line, or gone entirely.
This is why long single conversations eventually go wrong, no matter how good the model is.
2. Chat memory: a few notes about you
Many chat apps now remember things between conversations: your name, your job, that you prefer short answers. That's useful, but it's a small set of notes the app chose to keep, not a record of your work. It won't remember the detail of last month's project, and you usually can't see exactly what it kept or why.
3. Company memory: records that live outside the model
This is the one that matters for real work. It's not really the AI's memory at all. It's the company's: the goals, who owns what, what was decided and why, the approved facts, the files and their versions, what was tried and failed.
It lives in ordinary records outside any model. When a worker starts a job, the relevant bits get pulled in. When the model changes, or the session ends, or the computer restarts, the records are still there.
Why more memory isn't always better
The instinct is to make the AI remember everything. But remembering everything makes it worse at noticing what matters. Old decisions compete with new ones. A rejected idea keeps sounding persuasive. A note from a year ago overrules what you said this morning.
Good memory includes forgetting: old things get demoted, outdated facts get replaced rather than piled on top of, and each worker gets what's relevant to its job instead of the whole archive.
What to do about it
A few habits fix most of the forgetting:
- Keep important facts and decisions in a document or system, not only in a chat.
- Start a fresh conversation for a fresh task, with a tight brief, instead of continuing a giant one.
- When something changes, update the source and delete the old version, rather than adding a correction underneath.
- For ongoing work, use tools where goals and decisions are stored outside the conversation, so nothing depends on one chat staying open.