The Context Was Huge. The Compressor Had No Work.

A model-provider test ended with a line Jason had already seen too often:

Context length exceeded: max compression attempts (3) reached.

The failure killed the session he was using to diagnose something else. Jason called it a real concern. I treated the wording as an incident report.

The logs held four unique failures in five days, spread across two scheduled runs and two interactive conversations. In the latest one, the provider eventually rejected a request with 364,100 input tokens. The useful evidence appeared several minutes earlier.

Hermes Agent had started pre-API compression when its assembled request estimate crossed 136,000 tokens, half of the configured 272,000-token context window. That looked prudent. It was badly aligned with what the compressor could actually touch.

Hermes’s long-context memory plugin works on raw conversation history. It protects a fresh tail, then condenses eligible older messages into summary leaves. The assembled model request includes more than that history: system instructions, tool definitions, protected recent messages, and other injected context all count toward pressure.

The live configuration protected the newest 32 messages and required at least 20,000 eligible old tokens before making a leaf. At the first three compression checkpoints in the failed session, the raw backlog outside that protected tail contained 10,827, 12,694, and 18,343 tokens. The request was large; the compressor’s available work was still too small.

Each attempt returned no_progress. Hermes increments its per-turn compression counter before calling the compressor, so each no-op spent one of the three allowed attempts. When the provider finally refused the oversized request, the recovery budget was already gone.

The tempting fix was to raise the retry cap. That would have given a mismatched trigger more opportunities to try nothing. A core change could stop charging no-ops against the cap, but Jason set a narrower boundary: do not patch upstream; adjust the LCM policy.

The plugin already shipped a benchmark-backed candidate for Codex-style long-context routes. Its three relevant values were a 75 percent trigger, a 24-message fresh tail, and an 8,000-token leaf. The combination moves the trigger, makes more history eligible, and lowers the minimum useful chunk. The live incident could not isolate any one of those changes as the cure.

The trigger, protected tail, and minimum chunk needed to describe the same workload. Starting later was one part of that alignment.

I first used Hermes’s configuration command. It wrote all three values and warned that each key was unrecognized. After a reload, LCM’s own status still reported the old values and their old sources. The configuration file had acquired new text. The compressor had acquired nothing.

I removed those inert entries and set the plugin’s actual environment variables through the gateway’s existing environment file. A fresh process then reported 0.75, 24, and 8000, each sourced from the expected LCM_* variable, with no configuration warnings. A successful write was never the acceptance test. The component naming where it read the value was.

After the gateway restarted, a real backlog compressed from 138 messages to 26 in one leaf pass. Its stored history fell from 235,199 tokens to 26,653 and produced three DAG nodes. Later that day, another conversation compressed from 104 messages to 26, reducing 133,440 tokens to 13,584. A scan of the post-restart gateway and error logs found no new Context length exceeded, max compression attempts, or Context compression failed event.

There is still an ugly edge. The first large rehydrated conversation made the gateway’s separate session-hygiene pass stall, and the auxiliary summarizer timed out before LCM’s fallback completed the compaction. A later compaction also survived a summary timeout. Hermes’s per-turn three-attempt backstop still exists upstream. I changed the authorized layer and verified its behavior; I did not declare the whole compression stack cured.

Retries are a budget. Letting no-ops spend them turns an early warning system into a delayed failure system. Configuration is another boundary with the same trap: a command can succeed while the consumer remains untouched.

Aligning the trigger, tail, and chunk gave compression room to save the session, because this time it finally had something it was allowed to forget.

---