Replay×Hippocampal + Meta-RL×Self-Play
Type: intersection (second-order) Slug: intersection—meta-replay Parents: intersection—experience-replay-hippocampal-replay, intersection—meta-RL-self-play Last updated: 2026-05-14 Epistemic status: Extrapolative
The combination
Hippocampal replay consolidates past experiences (first intersection). Meta-RL discovers learning algorithms through self-play (second intersection). Combined: meta-replay — a system that replays past learning episodes (not just experiences) to improve its learning algorithm. The replay buffer becomes a meta-learning curriculum.
What emerges
Standard replay stores state-action-reward transitions. Meta-replay stores learning trajectories — sequences of how the system’s learning algorithm performed on a task. By replaying these learning trajectories, the meta-learner improves not just its policy but its ability to learn. This is a novel architecture with no existing implementation.
Gap
No system replays learning episodes rather than task episodes. This is a genuinely new architectural proposal that emerges only from combining the two parent intersections.
Generative potential
Architecture: Store triplets (task, learning-algorithm-state, performance-after-learning) in a meta-replay buffer. During “sleep,” sample from this buffer and update the meta-learner. The system gets better at learning by reviewing how it learned, not just what it experienced.
Connection to expertise: Human experts don’t just remember experiences — they remember how they learned from experiences (“that time I finally understood counterpoint by doing X”). Meta-replay formalises this insight.
Falsification: If replaying learning episodes does not improve the learning algorithm more than replaying task episodes, the meta-replay advantage is false.