Or: When Memories Get Good -- The Default Path Without Theoretical Breakthroughs
Epistemic status: Fairly confident in the core thesis (context + memory can substitute for weight updates for most practical purposes). The RL training loop is a sketch, not a tested proposal. I haven't done a thorough literature review.
Suppose there are no major breakthroughs in continual learning -- that is, suppose we continue to struggle at using information gathered at runtime to update the weights of a given instance of an AI model. If you try to update the weights at runtime today, usually you end up with catastrophic forgetting, or you find you can only make very small updates with the tiny amount of useful data you have [1] .
So, if you can’t train a day’s worth of information into the model, how could you end up with something that functions as if it were learning on the job?
Long Context Lengths, High Quality Summaries, and Detailed Documentation [2] [3] .
It’s a straightforward idea, and basically done today, just not particularly well yet. Laying it out:
1. The model does some task. In doing so, it gathers a tonne of information, say, a dozen novels worth. It can fit all of this in its context at once.
2. The model finishes its task. In concluding, it writes:
1. Short notes that it should always remember (we’ll call these memories), for example “This company prefers me to communicate in German”, “Documentation is available in this folder path”, and;
2. Detailed documentation about everything it did and everything it learnt [4] . This can be quite verbose.
3. The memories are kept in the context window. The documentation is available on disk and can be accessed on demand.
That’s it.
Why Doesn’t This Work Now?
Firstly -- it kind of does. In my own software projects I maintain a concise Claude.md file (which gets passed to each new agent on spawn), as well as extensive documentation which the Claude.md points to (and which the Claudes can search at will)