What is AI memory, and why does your assistant forget you?

Ask ChatGPT to plan your week and it will do a decent job. Ask it again next Monday and you'll repeat yourself: your work hours, your kids' schedules, the gym days, the fact that you hate early meetings. Most AI assistants have no durable memory — every conversation starts from zero.

Context windows aren't memory

The confusion comes from the context window — the amount of text a model can consider at once. Within one conversation, the model "remembers" what you said because it's literally still reading it. Close the tab and it's gone. A bigger context window makes conversations longer, not the assistant smarter about you.

Real memory is different: it's a system that extracts what matters from your conversations, stores it outside any single chat, and brings the right pieces back when they're relevant — days or months later.

How persistent AI memory works

Modern memory systems do three things:

1. Extract — after a conversation, the important facts get pulled out: people, preferences, projects, decisions, dates. Not a transcript — a distillation.
2. Organise — facts link into a knowledge graph: you → work at → Acme, Leo → wears size 4 shoes. Connected facts answer questions transcripts can't.
3. Recall — when you ask something new, the system searches that graph and hands the model exactly the context it needs.

one-chat.app — same memory, any AI
GPTClaudeGeminimemory onswitching AIs — memory stays
Book somewhere nice for Mum's birthday dinner.
remembered · 3 things
She's vegetarian, loves being near the water, and her birthday is the 14th — here are three waterfront places with standout vegetarian menus, all free that evening.
Swap the AI above — it still knows all of this.

What to look for

If you're choosing an assistant that claims to have memory, four questions separate the real thing from a gimmick:

• Can you see it? Memory you can't inspect is memory you can't trust. You should be able to open a page and read every fact it holds about you.
• Can you correct it? People change jobs, move cities, get new dogs. Stale memory is worse than no memory.
• Can you delete it? Genuinely — erased, not archived.
• Does it follow you between models? If your memory is locked to one AI provider, it's a moat, not a feature. The best setup lets your memory travel to whichever model is best for the task.

That last one is the reason we built 1Chat the way we did: one memory, every model. Your assistant should know you — and you should own what it knows.