Tuning AI Compaction: From Crashes to Early Triggers

The article explores the author's experience in configuring AI compaction to trigger early, preventing resource exhaustion issues faced with their OpenClaw AI agent.

Details

The author had previously faced an issue with their OpenClaw AI agent, where the context window hit 119% of the model limit, causing compaction to fail repeatedly. In this article, they explore their efforts to configure the compaction settings to trigger early, before the system reaches a critical state. The author started by checking the maintenance logs and attempting to set a compaction threshold, but encountered a silent breaking change in the configuration key name. They then experimented with different compaction modes, finding that the default mode, which treats the threshold as a hint rather than a hard trigger, worked better for their interactive sessions. The two key settings that made a difference were 'maxHistoryShare', which limits the fraction of the context window that can be consumed by conversation history before triggering compaction, and 'reserveTokens', which explicitly reserves headroom for the compaction process itself. By setting these values, the author was able to ensure that compaction triggers early, with guaranteed token reserves, keeping the agent responsive. The author also shares some lessons learned, including the importance of verifying configuration keys across versions, reserving resources for recovery mechanisms, and proactively managing the context window to avoid resource exhaustion issues.