We Cured the Wrong Amnesia

Ask a memory system what you told it last week and it answers. Ask it what it figured out last week and it goes quiet. We have stopped noticing the difference, and that is the whole problem.

For years the complaint about these models was always the same: it forgets what I told it. You explained the constraint, the model understood, and three turns later it was gone. The fix arrived, and it was real. Retrieval. Give the model a store it can read from, and the facts come back. The visible amnesia is cured.

Two amnesias, not one

But there were always two, not one.

The first is forgetting what you were told. The second is forgetting what you worked out from it. They feel like the same failure because they surface in the same place: a system that doesn’t know something it should. They are not the same failure. One loses inputs. The other loses the thing you actually paid for: the conclusion, the synthesis, the understanding that took the whole conversation to build.

Call the first Fact Amnesia. The second, Insight Amnesia.

Retrieval only touched the first.

What happens on a fresh read

Watch what happens on a fresh read. The system pulls the relevant facts back into view, and then, right there in the moment, it works out whatever it needs from them. It combines two passages. It notices a tension. It reaches a conclusion. Then the turn ends, and the conclusion is gone. Next time it pulls the same facts and does the same work again, from the start, with no trace that it ever reached the answer before. The facts stay. The thinking does not. It gets made again and thrown away, every read, forever.

The forgetting you can’t see

Here is the part worth pausing on. A system that forgets out loud is at least honest. You can see the gap, you work around it, you don’t give it anything important to keep. A system that forgets while looking like it remembers is far more dangerous, because the costume hides the loss. The facts come back smoothly, with the feel of true recall, so it reads as memory. The synthesis quietly evaporates underneath, and nothing in the exchange tells you it left.

Why the field thinks it’s solved

This is also why the field believes the problem is mostly solved. The amnesia that shows up in a demo is Fact Amnesia, and Fact Amnesia is exactly what retrieval cured. So the demo passes. The problem gets declared solved. Meanwhile Insight Amnesia is invisible by design: you only notice missing synthesis if you remember having reached it, and the system, of course, does not. That success hides the gap it never closed.

The same loss, one clock faster

If you have watched a long session degrade, you already know this loss under another name. Mid-conversation the context fills, something compacts it, and the model keeps talking but no longer reasons from the chain it built an hour earlier. The reasoning is still there in summary, but the live thread connecting it is cut. That is Compaction Amnesia, and it is the same loss as read-time synthesis. The only difference is the clock. Compaction throws reasoning away within a session. Read-time synthesis throws it away across sessions. One loss, two timescales.

Where the fix has to live

Which tells you where the fix has to live. Not in retrieving better. You can make the fact store flawless and Insight Amnesia is untouched, because the loss was never that the facts went missing. The loss is that the conclusion never became a thing that could last. It lived in the reader, and the reader forgets.

So the move is to let the synthesis itself survive the read. The conclusion gets written down and kept, not created fresh each time but recalled like any other fact, ready to be built on rather than rebuilt. Output becomes input. And once a conclusion can be read back, the next conclusion stands on it instead of starting again from the same raw facts, which is the one thing a pile of retrievable notes can never do. The store stops being a transcript you re-read and starts being something that builds on itself.

The right question

Retrieval was a real cure. It just treated the amnesia we could see. The one we can’t see is still running, every read, under a disguise convincing enough that an entire field mistook it for recovery. The question worth asking of any memory system is not whether it remembers what you said. It is whether it remembers what it concluded, or whether it only performs remembering well enough that you stop checking.

And if it holds its conclusions, and builds on them, and changes because of them, then “memory” is already the wrong word for what you are looking at. Which is the question worth opening next.