Memory/skill entropy: project-specific tactical data silently degrades agent quality over time

Описание и применение

Добавление простого аудита памяти в Hermes Agent позволяет выявлять и предупреждать о проектно‑специфических данных, которые ухудшают качество решений агента.

Задача

After weeks of daily use, the agent’s decisions degrade because project‑specific tactics accumulate in memory and are applied to unrelated projects, creating "experience debt."

Результат

Adding a simple audit function reduces memory noise, preventing project‑specific tactics from leaking into unrelated contexts and improving agent performance.

Рабочий процесс

• 1. Define STALE_PATTERNS regex list for project‑specific data.
• 2. Implement memory_audit(entry) to return warnings if patterns match.
• 3. Hook memory_audit into the memory write path; if warnings are returned, add a warning to the tool response.
• 4. Optionally flag entries older than N writes for review during read.
• 5. Run the audit regularly to keep memory clean.

Доказательства

Add a lightweight memory_audit function that runs on write and flags entries containing project-specific patterns

Источник

This issue is the automation of that exact cleanup.

Источник