Skip to content

Memory

MiniGrid's Memory environment - recall a cue seen at the start to pick the right exit.

See the environment class in this module for the task, layout and reward/termination details.

Bases: Environment

A genuine memory test (verified against MiniGrid's actual memory.py): the agent starts in a small room seeing one green object (Key or Ball, chosen at random each episode), walks down a hallway that ends in a T-split, and must walk onto the cell adjacent to whichever of the two end objects matches the one it saw at the start - the other one ends the episode with 0 reward. Unlike every other navix environment, this one is not solvable from the current observation alone once the start room scrolls out of view - it requires actually carrying information across steps.

state.mission stores only the success position - failure_pos is never a separate field, since it's always the mirror image of success_pos across the hallway's centre row (height // 2), and is derived algebraically in events.on_memory_failure instead (the same choice PutNear's move/target two-mission-target case made, for the same reason: avoid a second mission slot when the second position is always recoverable from the first).