hermes - 💡(How to fix) Fix feat: pre-turn memory health hook in run_conversation() [1 comments, 1 participants]

Problem

Standing rules in the system prompt that require proactive LLM behavior (e.g. "if memory exceeds 70%, compact entries") are consistently ignored. The LLM prioritizes the user's latest message over static rules, making these instructions unreliable regardless of how clearly they are written.

Root Cause

run_conversation() in run_agent.py (~L11713) has no pre-processing hook that checks memory health before each turn. The only mechanism is a text instruction in the system prompt, which depends on the LLM remembering to follow it — something it demonstrably fails at.

Proposed Fix

Add a pre-turn hook before the main LLM call in run_conversation() that:

1. Checks self._memory_store.get_usage_pct()
2. If > 70%, automatically runs compaction targeting < 60%
3. Logs the action for observability

Something like:

if self._memory_store and self._memory_store.get_usage_pct() > 70:
    compacted = self._memory_store.compact(target_pct=60)
    if compacted:
        logger.info("Pre-turn memory compaction: freed %d chars", compacted)

This ensures memory health is enforced mechanically, not by LLM compliance with a text rule.

Related

• memory-management-and-cleanup skill (user-created) documents this pattern
• Current workaround: cron job that polls every 6h (rejected as insufficient — the right fix is code-level)

Opened via feedback from a user session: standing rules in system prompt that require proactive LLM behavior are unreliable.