Consolidation policies and promotion thresholds between STM, LTM, and PM

Establish decision criteria and quantitative thresholds governing promotion of parameters and sub-networks from Short-Term Memory to Long-Term Memory and from Long-Term Memory to Permanent Memory within the Tri-Memory Continual Learning architecture, including upper thresholds informed by novelty detection, performance stabilization, and user engagement signals.

Background

The architecture separates learning across Short-Term, Long-Term, and Permanent Memory to address stability–plasticity trade-offs. Promotion of knowledge across these tiers is intended to safeguard mission-critical skills while enabling rapid adaptation.

The authors highlight uncertainty in how to design consolidation policies—specifically, when and how to promote learned representations across memory modules and which signals should drive these transitions. Formalizing promotion criteria is essential to prevent premature freezing of knowledge or excessive volatility.