Agreed. But, as an all-day agentic dev tool user, I know this pattern well. To avoid this regurgitated pixelation, I start a new conversation/chat as often as possible, and then human-in-the-middle curate the correct .md and json files when I kick off the next chat. My tools try their best to do exactly this, but they still suck.

These OpenClaw agents are not even close to there yet. They are throwing the entire thread into the context window each time, correct? This reminds me of what happens when you recursively upload a photo to an LLM, or upload a video recursively to YouTube. It gets not good. Compression sucks.

edit: holy crap, upon a bit of research, this appears to be the cause of AI sycophancy?

"Autoregressive amplification" or "context pollution"

https://arxiv.org/html/2601.04170