i asked
@claudeai
to do something, i open a new tab and type x . com.
claude reads a lot of files and asks for permission to edit a file(no im not one of those psychopaths who turn auto-edit on) but i'm still on twitter
what really happens if i accept it after 1 hour? what happens if i accept it after 1 day? what about 10 days? what if i never approve it?
claude whilst asking for permission is basically waiting to do a specific set of actions such as grep, command line execution, editing files.
if you come back within 1 hour, nothing has changed at all. not the chat history, not the tool call request, claude's memory of the conversation, the proposed diff / action parameters. all of it is still the same.
shouldn't the exact same thing happen if i come back after a day? here is why that's wrong.
you are assuming claude to be a stateful machine, like a paused program whose RAM is still alive when you return. however, that's not the case.
every time Claude produces a message, that completes a full forward pass through the model.
After that:
- computation ends
- KV cache is thrown away
- hidden states vanish
- no internal scratchpads persist
- no internal memory exists
there is nothing stored on Anthropic's servers except your prompts. the kv cache is also cleared regularly along with internal scratchpads and the conversation embeddings.
that's why if you come back after a long time, claude re-reads a lot of files to kind of start from scratch like it never read any files up until now.
so cool.