From Context to Plasticity: A Three-Timescale Architecture for Persistent Memory and Recursive Online Learning

Abstract

This work presents a theoretical architecture for continually adapting AI systems organized around three distinct timescales: (1) fast transformer key-value context for immediate interaction, (2) medium-term persistent recurrent memory with decaying importance and consolidation, and (3) slow, controlled parameter adaptation. The proposed pathway is context → persistent memory → importance estimation → consolidation → plasticity gating → bounded parameter update → changed future computation. The manuscript formalizes recurrent state dynamics, transformer attention, gated persistent memory, exponential importance decay, importance-biased retrieval, prediction loss, gated online updates, consolidation thresholds, and learned update rules. It also explains why adapting recurrent transition weights such as U changes state-transition dynamics, whereas updating memory changes stored state without necessarily changing the underlying transition rule. The package distinguishes inference-time state adaptation, memory adaptation, online parameter learning, meta-learning, recursive self-learning, and recursive self-improvement. It proposes controlled experiments comparing transformer, KV-context, persistent-memory, importance, consolidation, gated-update, and learned-update variants, with metrics for retention, transfer, forgetting, poisoning robustness, parameter drift, retrieval quality, and compute cost. This is a conceptual research architecture and falsifiable research hypothesis. It reports no experimental measurements and does not claim AGI, consciousness, human-level cognition, biological equivalence, or open-ended recursive self-improvement. The prefrontal-cortex comparison is functional only, referring to working-memory maintenance, goal persistence, error monitoring, and adaptive control. AI assistance disclosure: ChatGPT assisted with manuscript drafting, mathematical editing, structure, and package preparation. Final authorship and publication decisions remain with Chaman Prakash Kanth.