continual-learning

Prevent catastrophic forgetting during sequential model training with EWC, GEM, PackNet, and DARE.

Updated Apr 20, 2026

Search Tags:#continual learning #model merging #catastrophic forgetting #EWC #GEM A-GEM #LoRA adapters #MemRL

One-click install

npx skills add https://github.com/thistleknot/skills --skill continual-learning-thistleknot

Or copy as Structured Prompt for Agent▼

Please help me install this Agent Skill.
Skill: continual-learning
Source: https://github.com/thistleknot/skills/tree/main/continual-learning
Command: npx skills add https://github.com/thistleknot/skills --skill continual-learning-thistleknot

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

Continual learning prevents catastrophic forgetting when an AI model or agent is updated over time on new tasks, domains, or environments, degrading earlier capabilities.

Core Features & Use Cases

Stability mechanisms: Applies regularization-based methods like EWC to penalize changes to weights important for past tasks.
Replay-based learning: Uses episodic buffer approaches like GEM/A-GEM to constrain gradient updates so past-task loss does not increase.
Architecture and adapter isolation: Supports PackNet masking and LoRA adapter strategies (O-LoRA / InfLoRA) to isolate task-specific capacity.
Post-hoc consolidation: Performs adapter merging via DARE to combine multiple task specializations into a single model.
Distillation fallback: Uses LwF to preserve behavior via soft-label teacher predictions without storing old data.
Non-parametric agent memory: Integrates MemRL to update agent capability without modifying backbone weights, using retrieval plus Q-value–based memory updates.

Common use case: you fine-tune an LLM for new customer support domains while preserving earlier general skills, or you train a robot/agent across multiple environments while keeping performance from prior worlds.

Quick Start

Use continual-learning to update a model across tasks without degrading earlier competencies by selecting the method that matches your constraints (weights access, replay availability, and whether you use adapters).

Frequently Asked Questions about continual-learning

High-intent search queries and answers about installing and using this skill.

FAQPage Schema

How do I prevent catastrophic forgetting when fine-tuning an LLM on new domains?▼

You can prevent catastrophic forgetting in continual reinforcement learning by using MemRL, which updates agent capabilities via retrieval and Q-value memory without modifying backbone weights, or by applying GEM/A-GEM gradient constraints.

What is the best way to merge multiple fine-tuned LoRA adapters without losing task performance?▼

The best way to merge multiple fine-tuned LoRA adapters while preserving task performance is using DARE merging, a post-hoc consolidation method that combines multiple task specializations into a single model.

Can I preserve model performance on old tasks without storing previous training data?▼

Yes, you can preserve model performance on old tasks without storing previous data by using Learning without Forgetting (LwF), which distills knowledge via soft-label teacher predictions to maintain earlier behavior.

How do I choose between EWC, GEM, and PackNet for sequential model training?▼

Choose between EWC, GEM, and PackNet based on constraints: EWC regularizes weight changes, GEM uses replay buffers to constrain gradients, and PackNet isolates task capacity through network masking.