Finally we thank the reviewers who encouraged us to test our predictions and you for reading! The double dactyl poem summary didn’t make it past the editor so you’ll find it in the preprint – or here. 9/9

If these memories combine reusable codes, these cells should generalise their response when the cheese moves. We find cells like that too: if the cheese moves, the firing field moves accordingly. 7/9

Looks like it is! This cell has a new field in the bottom left after a replay. That’s the exact same location of this cell’s spike within the replayed trajectory. If you calculate the ratemap change, it peaks at the replay spike. 6/9

Or even better: hippocampus encodes those memories in replay. That builds a map for getting to cheese in locations without even going there, so they can be retrieved at that location later. These are memories of the future! 4/9

Crucially, these maps don’t need to be learned from scratch, but can be constructed from reusable grid (“you’re here”) and object-vector (“cheese-over-there”) codes. Hippocampus combines them in memory for the current environment. 3/9

If I’m a mouse that just discovered cheese, I’d like to know how to return to that cheese later. Instead of learning a plan, we propose building a map that comes with a plan for free. That solves planning in representation instead of computation. 2/9