Show notes
Cognitive scientist Dr Tomasz Wiśniewski on why human memory is mostly about forgetting, what that means for AI assistants that remember everything, and the design questions nobody is asking yet.
On this page
AI assistants are getting better at remembering things about us: our preferences, our projects, the names of our children and the way we like our emails written. That sounds like an unambiguous improvement. Dr Tomasz Wiśniewski isn't so sure.
Tomasz is a cognitive scientist who studies how human memory works — and, crucially, how and why it fails. In this conversation, he explains why forgetting is not a bug in the human brain but one of its most important features, and what designers of AI systems might learn from that.
In this episode
- 00:00 — Why we forget where we put our keys, and why that's useful
- 06:30 — Memory as reconstruction, not recording
- 14:10 — How AI assistants "remember" today
- 23:45 — The problem with perfect recall
- 33:00 — Consent, control and the right to be forgotten by your software
- 42:20 — What a well-designed memory feature might look like
- 49:10 — Listener question: should I trust my assistant's memory?
Memory is reconstruction
The first thing Tomasz wants listeners to understand is that human memory does not work like a video recording. When we remember something, we don't play back a stored file. We rebuild the memory each time, from fragments, guided by what we know and expect. That's why memories shift over time, why two people remember the same event differently, and why we can be completely confident about something that never happened.
"It sounds like a flaw," he says. "But it's what allows us to generalise. If you remembered every detail of every dog you'd ever seen, you'd struggle to recognise a new one. Forgetting the specifics is how you learn the category."
"A memory system that never forgets isn't more human. It's less."
How assistants remember now
Daniel walks through, in broad strokes, how many current AI assistants handle memory. A model has a limited "context window" — the amount of text it can consider at once. Some systems extend that by storing notes or summaries about a user and retrieving relevant pieces when they seem useful. Others simply let users save preferences explicitly.
Tomasz finds these approaches fascinating, and in many ways sensible. His concern is less about the technology than about the assumptions built into it — in particular, the assumption that remembering more is always better.
The trouble with perfect recall
People change. A preference stated once, years ago, may no longer be true. A difficult period in someone's life may be something they would rather not be reminded of. A detail mentioned in passing may be given far more weight than it deserves.
Human relationships handle this through a kind of graceful forgetting. A good friend remembers the important things, lets the trivial ones fade, and knows not to bring up painful subjects uninvited. Tomasz argues that software with memory needs to be designed with similar tact.
What good design might look like
Towards the end of the episode, Maya asks Tomasz to imagine his ideal memory feature. His answer has three parts:
- Visible. You should always be able to see what the system remembers about you, in plain language.
- Editable. You should be able to correct, update or delete any memory, easily and permanently.
- Decaying. Memories that are rarely used should gradually lose weight, the way they do for us, unless you choose to keep them.
"The best memory," he says, "is one that knows the difference between what matters to you and what you once happened to say."
Key takeaways
- Human memory reconstructs. We rebuild memories each time we recall them, which is why they shift — and why we can generalise from experience.
- Forgetting is a feature. Letting details fade helps us learn categories, move on from difficult periods and keep what matters in focus.
- More memory isn't automatically better. An assistant that recalls everything may surface outdated preferences or painful context at the wrong moment.
- Good memory is visible. People should be able to see, in plain language, what a system remembers about them.
- Good memory is editable. Correcting or deleting a memory should be easy, permanent and obvious.
- Good memory decays. Rarely used memories could lose weight over time, the way ours do, unless the user chooses to keep them.
- The real test is whether a tool leaves you more capable or less — whether it helps you think, rather than thinking for you.
Transcript excerpt
Daniel: Is there a risk that we start trusting these systems to remember things for us, and stop remembering them ourselves?
Tomasz: There's always been that risk. People worried about it with writing, with calculators, with phones. Usually we adapt. What matters is whether the tool leaves you more capable or less. A good notebook makes you think better. A good assistant should too.
Mentioned in this episode
- The reconstructive nature of human memory
- Context windows and retrieval in AI systems
- Data protection principles around access, correction and deletion of personal data
About our guest
Dr Tomasz Wiśniewski is a cognitive scientist whose research focuses on how people form, distort and forget memories in everyday life. He has taught undergraduate and postgraduate courses on human memory for over a decade and has advised design teams on how people interact with systems that remember. He is currently writing a book for general readers about forgetting, provisionally titled The Useful Fade. He lives with his partner, two children and an elderly cat who, he says, has perfect recall for dinner times.
Credits
Signal is hosted by Maya Chen and Daniel Osei. Our guest appears in a personal capacity.