Fixing the same agent mistake four times led to a shared memory file
The author found that correcting a GPT-6 Astra agent in chat only fixed that run, so they moved the correction into a small shared memory file that each new agent reads before starting the task.
my GPT-6 Astra agents kept making the same mistake even after I fixed it 4 times each correction worked perfectly for that run then a fresh session started and the exact same problem came back wrong filename wrong folder same correction again I eventually realized I wasn’t teaching the system anything I was teaching a conversation so I took the correction out of chat and turned it into a tiny shared memory file the next agent read it before touching the task then another agent did then the next run did too I stopped repeating the correction the mistake disappeared with it that changed what I save from agent runs not every conversation deserves memory but every correction you never want to make twice probably does an agent that learns inside one conversation is useful an agent system that keeps the lesson after the conversation dies is leverage
This post quotes another post: x.com/lummox_eth/status/2100288151897981255