Better agent memory

Agents spend significant amounts of time re-understanding projects because they do not have persistent memory. This means tasks take longer, context windows are smaller, and agents don't get better over time.

Current approaches for memory involve maintaining a knowledge base that the agent can continuously update and read from (e.g. a set of Markdown files or a database).

Managing memory is not something an agent user wants to think about. It's annoying. It should be part of agent infrastructure and harnessing.

Great agent memory should be:

Automatic: I don't need to ask it to remember.
Fast: I don't notice it.
Error-correcting: I don't need to tell it to forget X and remember Y instead.
This last bit is important. Users don't need photographic memory all the time. Forgetting outdated information can be useful.

Is there an architecture that enables the models themselves to learn on demand? Imagine a personalized model per user. The more I talk to a model, the more it resides in the latent space that I prefer. That could be interesting as a way to solve the problem outside the agent harness.