I want frontier AI providers to sell governed enterprise memory, rather than just longer context windows and short-lived prompt caching.

A company should be able to ingest a large body of knowledge once, turn it into a reusable memory or adapter, and attach it to future calls. I’m not describing full fine-tuning or reinforcement learning. I’m imagining something persistent, versioned, and governed that sits alongside the model.

Charge for ingestion, storage, refreshes, and usage. Give customers control over source lineage, permissions, retention, rollback, audit logs, and evaluations.

The shared responsibility model from cloud computing is a useful way to think about it: the provider secures the model and memory infrastructure, while the customer governs the knowledge attached to it.

That would be more valuable to me than paying large token costs to send the same context over and over. Longer context is useful, but it doesn’t answer the whole enterprise knowledge problem.