The Model Does Not Need Memory. The Situation Does.

I think I was asking the wrong question.

For a while, the question was simple: does memory make agents smarter?

It sounds like the right question. It is also a trap, because it assumes that memory should be judged as a generic intelligence booster. You add a memory layer, the agent remembers more things, and somehow the output should become better. More complete. More accurate. More human. More whatever word we are currently using to avoid saying “I hope this expensive thing works.”

After running more experiments, I think that framing is wrong.

Memory does not make the model smarter in any general sense. Most of the time, it cannot. The model already has a massive amount of general procedural and domain knowledge compressed into its weights. If your memory layer recalls information the model already knows, you are not adding intelligence. You are just adding a second path to say the same thing with more latency.