NPMCL: A Mechanistic Framework for Non-Parametric Continual Learning through Meta-Ability Cultivation

Abstract

Parametric update methods for Large Language Models (LLMs) in continual learning often face challenges such as catastrophic forgetting and the stability-plasticity dilemma. In this work, we characterize Non-Parametric Meta Continual Learning (NPMCL) as a structured approach that enables knowledge updates without subsequent retraining after one‐time meta‐training. This framework models adaptation as a Knowledge Compression-Decompression process, formalized through four core meta-abilities: (1) Query Generation for identifying information gaps; (2) Structural Matching for precise referential and temporal alignment; (3) Distillative Compression for extracting logical invariants from raw data; and (4) Constrained Reasoning for memory-guided reasoning and prior suppression.

We propose that these meta-abilities constitute a domain-agnostic cognitive pipeline, potentially allowing LLMs to adapt to dynamically changing environments by leveraging dynamic external memory. This work aims to formalize the mechanistic underpinnings of such meta-cognitive protocols. The proposed framework is informed by preliminary empirical observations from logic-aligned memory architectures (e.g., CoG-MeM), with Constrained Reasoning further explored in subsequent work (e.g., Knowledge-Constrained Reasoner). In this paper, we systematize the NPMCL paradigm, explore its potential for enabling continual learning by decoupling knowledge from knowledge-utilization capabilities at the LLM level, and incorporate a small-scale evaluation with knowledge data organized in different logical chain formats to provide an exploratory validation of the framework.