Continual Learning

How LLMs can learn continuously from experience, adapting their behavior over time without full retraining. The bridge between static model weights and dynamic agent memory.

The Problem

Pre-trained LLMs are frozen after training. Current “memory” approaches (structured memory + in-context learning) face scaling limits:

  • VRAM is expensive
  • Context length requires rare long-range dependency data
  • Self-attention has quadratic complexity

Approaches (2x2 Framework)

From Predictions on Continual Learning:

Low CoverageHigh Coverage
NarrowIn-context learningLoRA fine-tuning
WideStructured memoryFull fine-tuning

The “missing piece” is generic test-time LoRA fine-tuning for narrow, high-coverage cases (e.g., personalizing for a specific user or task).

Current State

  • In-context learning: Agents (Claude Code) and memory frameworks (Mem0) handle well
  • Full fine-tuning: Cursor’s continuous RL releases new checkpoints every 5 hours
  • LoRA at test-time: No mature technique yet — predicted to emerge from big labs by end of 2026
  • Feature-guided LoRA: Using SAE features to make test-time adaptation efficient and interpretable
  • Doc-to-LoRA — documents → LoRA adapters in <1s
  • Titans paper — smart optimizer on fixed-size memory module
  • On-policy distillation with LoRA matches full SFT at 9x lower cost

Open Questions

  • How to weight experiences? (garbage in, garbage out)
  • Self-generated token datasets are small and noisy
  • How to prevent catastrophic forgetting during continual adaptation?
  • Can SAE features provide stable, interpretable directions for adaptation?

Connections