Capability to Reliability to Learning (Part I)

First it works, then it sticks, then it learns...

It's almost 2026, and yet the agents still have amnesia. A single agent can follow complex instructions, use tools, and work autonomously for longer with each iterative progress.

However, when asked the same question again, the model starts from ground zero. The model doesn't correctly remember what worked, what failed, or what they figured out along the way. We are starting to see the capability to turn out more reliably than before, but we haven't quite figured out how to make these models learn.

Memory is a problem

We can break these down to three layers: session layer, user layer, and learned layer. We have started with the primitive version of these using "Skills" and "RAG". Skills, in many ways, can be viewed as "Retrieval Augmented Capability," where the LLMs now have a relatively reliable way to use a "skill" on demand.

The session layer can be simply viewed as the context. This is the most basic form of memory, where storing and reused in context.

The user layer is about specific facts about a user that persist for as long as the sessions are alive.

The learned layer is where, for humans, the knowledge compounds happen despite having a very small set of data compared to the LLM pre-training data. In many ways, human brains are a lot more sample efficient compared to the current transformer based LLMs.

We still haven't quite figured out how an agent interacts with the real world and how it discovers insights that can apply to more than one user. In this specific example, we can view the "learning" to occur at the retrieval and not within the existing model weights. This in theory, can lead to continuous system improvement without fine-tuning or retraining. A system where the knowledge base grows with every interaction.

More to come on this topic. This is just the beginning portion of the hypothesis.