How do you keep long sessions from eating the whole context window?

I've been running multi-hour OpenClaw sessions and the context window fill-up is my main pain. Native compaction kicks in late (around the threshold) and it's all-or-nothing — once it summarizes, older detail is gone. What I wanted instead: compress \*gradually\*, every turn, but keep the last few turns completely raw so the agent doesn't lose the thread it's mid-way through. I ended up writing a Plugin SDK hook on before\_prompt\_build that does this — folds older turns into a compressed episodic view, keeps the trailing turns verbatim. On a long session it cut the re-sent context by roughly 80% without the agent losing track of earlier turns. Two questions for people running long sessions: 1. Do you rely on native compaction, or roll your own context management? 2. Has anyone found the right "keep N turns raw" number? I'm defaultingto 4 but it feels workload-dependent. (If useful, the hook is here — MIT core: [https://github.com/compresh/compresh-mcp](https://github.com/compresh/compresh-mcp) — but mostly curious how others are handling this.)