Prosaic Continual Learning
New essay argues long context and better memory systems can bypass the need for runtime weight training.
A new concept called 'Prosaic Continual Learning' is gaining traction in AI circles, proposed by researcher HunterJay in a detailed LessWrong essay. The core thesis addresses a fundamental roadblock: the difficulty of updating a neural network's weights at runtime without causing catastrophic forgetting or requiring massive datasets. HunterJay argues that instead of chasing theoretical breakthroughs in weight-based continual learning, a more practical path exists. By leveraging the rapidly expanding context windows of models like Claude 3.5 Sonnet (200K tokens) and GPT-4o (128K tokens), AIs can write and store high-quality summaries, 'memories,' and detailed documentation of their experiences. These notes are then made available to future instances of the model, creating a functional form of learning without ever modifying the underlying model parameters.
The essay details how this approach is already partially implemented in systems like Anthropic's Claude memory features and user-maintained 'Claude.md' files, but is limited by current models' abilities. The key challenge is that writing useful, concise memories requires a strong 'theory of mind'—the AI must understand what a fresh copy of itself wouldn't know and how that copy would use the information. HunterJay notes that early memory systems in ChatGPT often created irrelevant notes, and models like Kimi's K2.5 still struggle with over-attending to context-window memories. However, the prediction is clear: as models become more capable at retrieval, summarization, and understanding their own knowledge states, this prosaic method will become a powerful, default form of continual learning by 2026-2027, effectively sidestepping one of the field's hardest technical problems.
- Proposes using long context (e.g., 128K-200K tokens) and AI-written memory systems instead of runtime weight updates to avoid catastrophic forgetting.
- Identifies current limitations: models need better 'theory of mind' to write useful memories and avoid over-attending to irrelevant context notes.
- Predicts 'prosaic' memory-based learning will become a standard, powerful approach by 2026-2027 as model capabilities in retrieval and summarization improve.
Why It Matters
Offers a practical, near-term path for AI agents to learn from experience without solving the hard problem of continual weight training.