mllun

Layered long-term memory for DeepSeek Harness: a ~600-character hot layer that evicts its own overflow, month-archived cold entries with usage-ranked keyword recall, and extractive day/week/month summaries that keep injected context small.