There are two ways to make an agent improve.

One is to leave the model unchanged and accumulate documents, embeddings, summaries, skills, and longer histories.

The other is to let experience modify the model.

Boltzbit’s Bayesian Self-learning Transformers (BAST) investigate the second path by converting live data into targeted parameter updates. Its theoretical analysis estimates roughly 1,000× lower compute requirements than conventional training.

The economics matter. External memory makes future requests retrieve and process more state. Weight adaptation moves useful knowledge into the model itself.

The next benchmark is whether those updates remain stable, selective, and generalize beyond the interaction that produced them.

If they do, continual learning becomes part of the architecture, not another agent feature.