AI remembers. But which decision still applies?

I changed the examples on a website several times. The main page was already telling the new story, but pieces of the old one remained in the corners: a number that no longer meant anything, an explanation pointing to content that had been deleted, and copy that was no longer used on the page but still existed in the code.

Each sentence could read perfectly well on its own. Back on the page, the facts no longer lined up.

Calling these missed edits is fair. But a more useful question is: what was the reference point at the time? If a review covers only the places that AI and I remember from the conversation, old content outside that remembered set may stay there indefinitely.

Nothing had been lost. What had broken was its relationship to the current decision.

Saved does not mean still valid

What I actually found was leftover copy, not the result of a model memory test. But it made me reconsider the context I give AI.

Suppose an old plan says to use example A, and a later plan switches to example B. Both are saved and searchable. Without an explicit replacement relationship, the next task still has to resolve the ambiguity: did the decision change, or do different pages use different examples? Is the new plan a confirmed decision, or a proposal that has not yet been adopted?

Saving another file does not answer those questions.

So I now care less about whether AI can find something said in the past than whether, having found it, AI can tell what that statement should still govern today.

Give current decisions a clear starting point

Later, I put together a canonical fact document for the website. Rather than collecting all the copy, it specifies which example is in use, what the numbers mean, and which facts every page must be built around.

That distinction matters. Copy can take many forms. The facts should not depend on each page independently remembering them.

If entire passages are copied into multiple documents, the next change requires maintaining multiple copies. The canonical document should state the current facts; other places should express them. Old versions belong in the history, where they explain how a decision came about rather than continuing to serve as another current plan.

But writing that document solves only half the problem. The homepage, detail pages, and design notes may all retell the same thing. When a decision changes, those places still need to be checked for old assumptions carried into new wording.

Writing down a new decision and taking the old decision out of use are not the same thing.

The newest date does not automatically win

The simplest approach might seem to be: just read the latest record.

But an experimental proposal written today does not overturn a confirmed plan from yesterday merely because its date is newer. A new rule may also apply only to one project; it should not casually override other contexts.

What needs to be explicit is quite ordinary: where the decision applies, whether it has been confirmed, when it takes effect, and what it replaces.

For conflicting decisions covering the same scope and period, there should be a clear current reference. Rules with different scopes need explicit boundaries, not a forced merger under the instruction to use whichever is newest.

This does not mean adding a set of fields to every passing note. When something will affect the next action, even a clear sentence can help:

This plan has been replaced by the new version and is retained for historical reference only. For the current implementation, start with the new version.

The point is not the format. It is that the next person or AI using the material should not have to guess again.

Keep history without letting it pass for the present

I do not want to delete every old plan. Without them, it is hard to explain why a path was abandoned, or to judge whether a later problem arose from an earlier trade-off.

But keeping history and acting on history are different uses. Reviewing a process requires seeing what changed; continuing the work requires knowing the current basis for action. If retrieval brings back an old plan, its replacement notice and the link to the new version should come with it—not just the answer that was once correct.

Neither a canonical document nor a replacement notice guarantees that AI will never use the wrong material. Whether they work still has to be checked in the next real task: what basis did AI use, and did it bring discarded content back into the result? Adding a marker alone does not mean the problem is solved.

What I want a memory system to retain is not just what I once said, but what changed afterward.

Remembering the past should help us avoid retracing our steps. It should not let every past decision keep deciding the present.