Anti-LLM Sentiment Considered Harmful

How do you bypass context compaction and sampling? How does that not burst the context window of the LLM if you constantly keep everything in memory? Edit: I’m also a bit suspicious about any claim of an LLM having “senior developer” capabilities, considering that the statistical nature of this tool prevents it from thinking outside the box. It can only take statistically relevant samples of what’s inside the box.