Feed
Streaming Recurrent Memory
A memory that learns from live data streams by fixing what it writes and learning only how it reads.
0.99
success rate
T-Maze success, 8 key-value pairs
Delayed-reward memory task solved without replaying past experience
How it works
No replay, no history
Streaming learners update once per input and cannot backpropagate into the past. So later feedback cannot fix how memories were written.
Fix write, learn read
SRM stores each input under fixed random keys mixing content and a clock. It learns only queries, decay rates and a tracking channel.
Beats recurrent baselines
With one-step backprop, SRM wins on associative recall, byte-level TinyStories loss, and delayed-reward T-Maze control.
From the paper