An LSTM is a Conveyor Belt of Memory

Added: 2026-06-05

[00:00:00]An LSTM keeps a thread of memory running straight through time. Picture a conveyor belt carrying the cell state from one step to the next, almost untouched.

[00:00:11]What controls that belt are three little gates, and each one is really just a sigmoid that squashes everything into a number between zero and one. A valve that's open, closed, or somewhere in between.

[00:00:27]The forget gate decides how much of the old memory to erase, while the input gate decides how much fresh information to write. So, the new cell state is ft times the old state plus it times a candidate.

[00:00:44]And finally, the output gate reads from that memory, passing it through tanh to produce the hidden state ht that actually leaves the cell. So, an LSTM isn't magic. It's a conveyor belt of memory with three valves choosing what to keep, what to add, and what to reveal.