Continual Learning in Token Space Fresh

Published: December 11, 2025 | Original Post

---

Updates to learned context, not weights, should be the primary mechanism for LLM agents to learn from experience.

The Learning Gap

While humans continuously improve through experience, current AI systems fail to meaningfully improve over time. Modern agents are "weights plus context" -- their behavior emerges from parameters combined with prompts, documents, tools, and conversation history.

Why Weight-Space Learning Falls Short

1. Training incompatibility -- Methods requiring careful data curation do not scale to continuous online updates
2. Deployment challenges -- Managing millions of users creates privacy concerns when mixing personal data into shared weights
3. Unanswered questions -- Unclear learning signals and difficulty weighting recent vs. older information

Token-Space Learning Advantages

• Interpretability -- Human-readable memories enable direct inspection and debugging
• Portability -- Context transfers across agents, providers, and model generations without weight lock-in
• Control -- Forgetting is trivial, rollbacks are simple, versioning works like text files

Formal Framework

An LLM agent is defined as the pair (theta, C), where theta represents model weights and C represents the context window. Token-space learning optimizes C instead of theta, making catastrophic forgetting prevention trivial -- simply restore previous versions.

Key Insight

Agents that can carry their memories across model generations will outlast any single foundation model. The weights are temporary; the learned context is what persists.