Learning without weights: where continual learning actually lands

Position, signed by Mei Tanaka, 30 June 2026. This is the lab's view on the continuous-learning field: whether models will learn from use in a way that changes their weights, on what timeline, and what to do in the meantime. It draws on two validation runs of a next-horizon field, on the memory and personal-wiki experiments (x-memory-bench, x-personal-wiki), and on the early-stage x-continuous-learning experiment, which is testing the non-weight-bound route directly.

What is demonstrated. Learning outside the weights works and is in production. An agent with an episodic store, summarised recall and a consolidation step stops re-deriving decisions (c-agentic-memory-2); an agent whose experience is compiled into a persistent skill wiki improves on repeated task classes across sessions with no training run (c-continuous-learning-1). Both are retrieval and compilation, not learning in the weights, and both deliver most of what a client means when they ask for an agent that 'gets better'. Inside the weights, the demonstrated results are laboratory-scale: continual fine-tuning without catastrophic forgetting at small model sizes, on curated streams, with a human deciding what to learn (c-continuous-learning-2).

What is hype. 'Self-improving' agents. Every current example is either a retrieval loop relabelled or a training pipeline with a person in it. Also hype: the idea that continual weight learning is imminent because the frontier labs are hiring for it. They are hiring for it because it does not work yet, and the hiring signal is carried as an inference, not a fact. And it is hype that learning outside the weights is a stopgap. On the evidence, the stopgap has the better auditability, the better forgetting story under APP 11, and the better cost profile, and it may simply be the answer (c-continuous-learning-3).

What would have to be true for the position to be wrong. A demonstrated continual-learning result at production model scale, on an uncurated stream, without forgetting and without a human curating the updates — none of the four conditions has been shown together. If one of the labs shipped it, the substituting join from continuous-learning to agentic-memory would fire and most of the memory layer would become unnecessary. The field's depends-on edge records exactly that trigger. Our estimate is 4yr+, with low confidence on the far end.

What we would do. Invest in the non-weight route as the practice's continual-learning story: memory defaults, compiled skill wikis, consolidation policies, and the audit and forgetting controls around them. Do not sell 'self-improving agents'. Keep the weight-bound route as a watched next-horizon field with a quarterly read of the literature and no spend beyond that. Re-run the position annually or when x-continuous-learning concludes.

Moderate strength: the non-weight evidence is tested, but the timeline claim on the weight-bound route is a forecast about research nobody has published, and it should be read as such. The red team's strongest case — that a single architectural result could collapse the timeline — is fair and is why the dependency trigger is explicit.