Cadence's recall memory keeps a list of the stream's recent moments. Before each decision the present moment looks over that list and reads back a blend of the moments that match it. Until 0.89.0 a match was decided by content alone: what a moment looked like, plus a fixed nudge toward newer ones. Nothing said when a moment happened, so "the frame just before this one" and "a similar frame twenty moments ago" were hard to tell apart.
0.89.0 gives every memory slot an age, and each read head learns a bonus per age that depends on what it hears right now (recall_position, 32 ages with their own bonus, older ones grouped by doubling). The bonuses start at zero and move only through taught lessons. Rewards never move them, so a brain that learns only from points reads exactly as one without the feature.
Illustration, not measured values: the read's weight on each slot is its content match plus the learned bonus for its age.
MinAtar's Breakout (Young & Tian, 2019) is a 10×10 version of the 1976 Atari game, built as a small, fast reinforcement-learning benchmark. MinAtar normally adds a "trail" channel showing where the ball was one moment ago, because a single frame doesn't say which way the ball is moving. These players don't get the trail. They see one frame: ball column, ball row, paddle column. A ball rising to the left and a ball falling to the right look identical. Remembering the previous moment is the only way to know.
A coach that knows the ball's direction plays the ball forward on a copy of the game and says which way to move. Every player plays the game itself and hears the coach's move at every training moment. Cadence takes it as a lesson, then acts. The MLPs take one gradient step. Every 1,000 training moments each player sits a test of three games with no coaching. All players get the same test games.