The past few weeks I've been experimenting with using less context and less memory and it's been going really well. Where before I'd try to do a bunch of fairly related things in a single session, experimenting with compacting more or less frequently, now I'm clearing my context or exiting and restarting claude and codex. It seems to help it focus on the task at hand, hasn't tended to go off into the weeds as much, and my token costs have dropped way down.

Combined with a good AGENTS.md, it seems to be working really well.

That’s been my experience as well. I find I usually get better output if I create a new conversation for each thing I need. I’ve found that the only times it’s better to continue an existing conversation is if I want to have it make small improvements or changes to something it just wrote, as it tends to do better with the previous context still there. But even that only goes so far, then the scale tips and it works much better with a clean slate. I especially don’t want totally unrelated conversations polluting the context which is why I have all memory features turned off in all the web chat UI’s for the models I use.