when your AI's context window fills up and you hit compact, it squeezes your whole conversation into a short summary. whatever didn't make it into the summary or the AI's memory is gone.
so never compact before you save the things you can use later.
your AI does the saving. you just tell it what to save.
what compact actually does
every AI chat has a context window. that's how much of your conversation it can hold at once.
in Claude Code or Codex, when it gets full, /compact turns the conversation into a summary so you can keep going in the same chat.
the summary keeps the big picture. it can lose the small stuff: a decision you made way earlier in the chat, the reason you made it, or something your AI got wrong.
you can also type what you want kept right after /compact and it'll keep it.
when to compact
compact around 50 to 70% full. not at 95%, which is usually the default and automated.
or skip it: finish the task and roll over to a brand new chat instead.
[nothing gets compacted before your logs are in]
what "logs"? a log is the record of a session: every mistake, error, decision and piece of feedback.
here's how Auny uses them, as the worked example. she keeps logs of every session she has with her AIs. once a week she synthesizes them and teaches her AI what it can do better: new rules, parameters and guards that improve its default behavior and personalize it to how she works.
@suavecito585 was the one who first gave her the idea and taught her about keeping session logs.
the 4 things to save first
- everything you've worked on so far
- finished decisions go into your vault (the notes folder all your AIs read and write to) and your AI's memory
- whatever you're in the middle of goes into a plan doc
- all your feedback, every decision, and a log of where your AI messed up
the 4th one matters the most. keep it too.
why the f-ck ups matter
the logs are how your AI gets better at working with YOU.
when it messes up and you correct it, that correction goes in the log. then use the logs to teach it how to improve its behavior, how it thinks and what it does for you and with you.
over time it gets personalized to you and how you work. Personalized Intelligence is where it's at.
Auny's agent-personalizer sets everything up the way she has hers: grab it on Github or install it from your terminal.
npm i agent-personalizer
the more detailed the logs, the better your AI gets over time.
have your agent do it
you don't need to build anything for this. before your next compact, paste this into your chat:
before we compact: save everything we've worked on so far. put every finished decision in [your notes app or memory file]. put whatever we're in the middle of in a plan doc. then write a short log of my feedback, the decisions we made, and anything you got wrong and how i corrected it. tell me when it's all saved
then compact, and point it at the plan doc when you pick back up.
watch your context. when it hits 50 to 70%, paste the prompt above, check that everything saved, then compact.
you'll notice the difference the first time you don't have to re-explain yourself. and compacting a convo saves token usage, because your AI doesn't have to keep dragging the long string of the convo along.
1:1 call · matches this post
AI Workflows for Creators
want your AI saving its own logs and plan docs before every compact, so you never re-explain yourself? That's the AI Workflows for Creators call.
