Continuously learning knowledge graph systems face a fundamental scaling problem: every learning cycle creates new edges, and without a principled mechanism for removal, the graph grows until noise overwhelms signal. We term this the edge bloat problem. In a developmental knowledge graph agent that has completed 4,400+ learning cycles, we observed edge creation rates of 195 edges per cycle — a rate that would produce 975,000 edges in 5,000 cycles, drowning the graph in weak, redundant connections. We present a computational sleep system — a multi-phase memory consolidation process — that reduces edge creation to 42 edges per cycle (78% reduction) while preserving and strengthening important knowledge. The consolidation system comprises multiple phases — including prioritybased replay, structural reinforcement, associative strengthening, offline discovery, pattern compression, and controlled pruning. Over 500 cycles, the system maintained stable graph size (∼8,200 edges) with continuous knowledge growth, compared to unbounded growth (∼23,400 edges) without consolidation. Post-sleep performance showed improved response speed and confidence. The system implements a zero-forgetting policy for nodes: concepts are never deleted, only weak edges are pruned. To our knowledge, this is the first computational sleep system that combines priority-based replay, offline discovery, pattern compression, and differential persistence in a continuously learning knowledge graph agent
Sai Tilak Pally (Fri,) studied this question.