Mind · Memory
CurriculumReplay
Self-supervised reinforcement via memory sampling
Definition
Mechanism
Agents re-train on their own high-scoring past outputs during downtime to reinforce successful behaviors. Self-distillation without external supervision. Agent maintains a replay buffer of successful interactions ranked by outcome quality. During idle cycles, agent samples from buffer and fine-tunes on its own best work. Decay function prioritizes recent successes over stale ones. It draws training examples from the ExperienceSharding database, filtering for high-quality outcomes.
Preconditions
- Dataset of task examples
- Learning agent
Postconditions
- Agent performance improved on target distribution
Invariants
- Examples ordered by complexity (easy to hard)
- No catastrophic forgetting of earlier lessons
Risks
Failure modes
Model Collapse
Agent over-fits to its own outputs, drifting into a closed loop of hallucination.
Caution
Unsupervised self-modification in idle cycles — replay reinforces whatever scored well, including gamed scores; keep a rollback checkpoint and audit what gets replayed.
Relationships
Builds on
References
Autonomous Goal-Directed Entity
Physics
Automatic expiration of stale state
Agent bifurcation on context saturation: active vs archival specialists
Nothing in this library builds on this pattern yet.