Stop learning the same thingtwice.
A memory that lives inside Claude, ChatGPT, and your terminal. It notices where you struggle, keeps a map of what you truly remember, and checks in right before you forget.
You've asked this before. Somewhere in a thousand dead chats is a version of you who understood it.
A month later you ask again. The model explains again. It feels like learning again.
It isn't. Reading a good answer is not remembering it. The chat scrolls away, and the understanding goes with it.
The right time to make a flashcard is the moment you're struggling. You never notice it — you're busy struggling.
Claude notices. It's right there while you work something out. It sees which step you tripped on, which idea you reached for and missed, which question finally made it click.
So it plants a card at that exact spot — and moves on with you. You never write a card. You never open a deck. The review comes to you, inside whatever chat you're in, right before the memory slips.
ClaudeTuesdaycard planted · “why must two majorities overlap?”marked hard · quorum overlap · returns soonerYou studied. It watched where the effort went — and planted the cards you'd never have thought to write.
It knows what you know.
Quizzing is the surface. Underneath, latent memory keeps a living map: what's rooted, what's still taking hold, what keeps slipping, where the gaps are. Every session starts from that map — so the next question is always the one you need.
Not a deck. A map of your memory — read fresh at the start of every session.
Forgetting is not a character flaw. It's a curve.
Ebbinghaus plotted it in 1885: memory decays on a schedule. A quick recall at the right moment resets the curve — and each reset makes it flatter. The method has worked for a century. What never worked was making the cards and showing up. That's the part latent memory takes off your hands.
Each green dot is one question, asked inside a chat you were already having.
One memory. Every chat.
Study in Claude Code tonight, answer a card in ChatGPT while you wait for coffee, pick the thread back up on the web tomorrow. Your memory isn't in any one of them — they all just reach for it.
Claude Sunday, 9:40pm2 due · 1 slippingrooted · next in 3 weekssix days later, on your laptop — the session opens where you're weakest
faster this time · next in 12 dayson your phone, in line for coffee — a 20-second review, then back to your day
3 due · log replication firstgot it · next in 9 daysa study session in the terminal — same memory, wherever you already are
Claude Friday, on a walkgap closing · 2 cards plantedthinking out loud, speech-to-text — rambling counts as studying here
got it · next in 16 daysOpenClaw, in your messages — your assistant carries the same memory
2 cards linked · compaction · snapshotsa Hermes research agent — even your agents know what you're learning
1 due · election timeoutsgot it · next in 2 weeksopencode, end of day — a different terminal, the same thread
14 cards · 9 rooted · 1 gap leftpi, in a study session — it reads your map before you ask
Under the hood it's an MCP server — one small brain any AI client can plug into. You never think about it.
Latent memory can't learn anything for you. All it can see is effort — a wrong guess, a second try, the question that finally made something click. That's the raw material for cards.
A chat that's one question and a skimmed answer gives it nothing to work with. It'll sit quietly and plant nothing. That's by design.
The understanding has to be yours. Keeping it is the part that can be handed over.
I have ADHD. What finally worked for me was making AI ask me the questions until I could derive things myself. Two weeks later, even that was gone. So I'm building the missing half.
latentmemory is coming soon.
Leave an email and you'll hear when it opens — and help decide what it becomes.