Memory Is an Index. Reviewed Artifacts Are State.

I have started separating two ideas that are often collapsed into one word: memory.

One kind of memory helps an AI find relevant history. It recalls that a decision was discussed, locates the note where a preference appeared, or retrieves a prior attempt. This is useful, but it is still an index. It points back into a large and imperfect record.

The other kind changes what the system is expected to do next. A decision becomes a constraint. A preference becomes a design rule. A proposal becomes the current plan.

That second kind should not happen invisibly inside a conversation.

For consequential work, I want durable state to live in a reviewed artifact: a brief, a decision record, a context pack, a test fixture, a project-state file. The name matters less than the properties. It should be inspectable, versioned, attributable, and easy to correct. A human should be able to see what changed and decide whether that change deserves to persist.

This creates a useful boundary:

Memory retrieves. Review promotes.

A model may remember that I once preferred a particular approach. That does not mean the preference is still current, applies everywhere, or should silently become policy. Retrieval can surface the evidence. Promotion turns it into state.

The distinction adds friction, but it is productive friction. It prevents an old remark, a misunderstood sentence, or a speculative idea from quietly governing future work. It also makes forgetting less mysterious: correct the artifact, supersede it, or remove it. You do not have to negotiate with an invisible internal history.

Not every conversation needs this machinery. For low-stakes personal assistance, lightweight memory may be enough. But when an AI can affect a product, a team, a customer, or another person, I want the important state outside the conversation and inside something we can review together.

Memory is valuable because it helps us find the past. It should not, by itself, get to decide the future.