I reduced my OpenClaw token usage by 10x

A user successfully reduced their OpenClaw token usage by a factor of ten, dropping daily consumption from approximately 2.2 million to 256,000 tokens. This significant optimization was achieved not by merely reducing the agent's heartbeat frequency, but by diagnosing and rectifying issues related to uncompacted session histories, setting explicit compaction triggers, and eliminating redundant instruction processing between a daily cron job and the agent's heartbeat mechanism. The core problem identified was that persistent sessions were replaying their entire, growing transcripts on every call due to inactive compaction.

The key technical finding was that all persistent sessions had `compactionCount: 0`, meaning their message histories were never summarized, leading to high `cacheRead` token usage. The solution involved configuring `agents.defaults.compaction` with `maxActiveTranscriptBytes: "500kb"` and enabling `memoryFlush` with `forceFlushTranscriptBytes: "1mb"` in the `openclaw.json` configuration. Additionally, a critical duplication was discovered where a daily cron and the heartbeat both processed the same `HEARTBEAT.md` instructions, resulting in redundant evaluations and unnecessary token expenditure.

This analysis highlights critical considerations for OpenClaw developers building persistent, long-running agents, emphasizing the necessity of proactive memory management and explicit compaction strategies. It underscores that relying solely on the model's context window pressure for compaction can be insufficient, leading to silent token bleed in generous context environments. The findings inform best practices for designing cost-efficient and robust agent architectures within the OpenClaw ecosystem, particularly in multi-agent or cron-driven setups where instruction duplication can occur.

This signal is highly relevant for OpenClaw developers and operators focused on cost optimization and efficient agent deployment. Developers should meticulously review default compaction settings and actively monitor `cacheRead` token usage within `trajectory.jsonl` logs to identify similar inefficiencies in their own agent implementations. Operators managing OpenClaw deployments can leverage these insights to implement more robust memory management policies and audit agent configurations for hidden token costs, ensuring sustainable and scalable agent operations.