Context decay in AI isn't a model problem. Here's what it actually is.

Context decay is not the model getting worse; it is the session losing the shape of what matters.

At some point in an extended AI session, the outputs start to drift.

The model is still responding and the responses are still coherent, but something has shifted. The outputs are a little more generic, a little less tuned to the specific context you have been working in. The earlier precision has softened.

Most teams notice this and assume the model has hit a limit. They start a new session, the context resets, they re-explain everything, and the first few responses are sharp again.

This is context decay.

The model does not degrade. What degrades is the ratio of relevant context to total context. As a session extends, the earlier and more specific context gets progressively diluted by everything that comes after. The model still has access to a lot of material, but it attends to the important parts approximately rather than precisely.

The intuitive fix is to add more context at the start: a longer brief, more detailed instructions, more examples. This creates a different problem. The model receives more information than it can weight precisely, attends to all of it approximately, and produces outputs that are competent but generic.

More context does not solve the decay problem. It just moves it.

The solution is architectural. Instead of loading context into a session and hoping it holds, you build a structure that lets the model navigate to the relevant context for a specific task.

This is the core logic of SR-SI - Simulated Recall via Shallow Indexing. The index is shallow by design. The model retrieves what is relevant to the current task rather than processing everything that might be relevant to every task.

Context that is navigable does not decay the same way loaded context does. The relevant information stays accessible because it is indexed for retrieval, not buried in a session that has been running for three hours.

The model is not the main variable; the architecture is.