Six-layer memory architecture combining WAL protocol, vector search, git-based knowledge graphs, and cloud backup.

• Hot RAM (SESSION-STATE.md) with Write-Ahead Log protocol ensures context survives crashes; warm store (LanceDB vectors) enables semantic search across memories with auto-recall
• Cold store uses git-notes for structured decisions and learnings; curated MEMORY.md archive plus daily logs provide human-readable long-term context
• Optional SuperMemory cloud sync and Mem0 auto-extraction reduce tokens by 80% while deduplicating facts across sessions
• Strict memory hygiene workflow: write state before responding, store decisions silently, archive weekly, and clear irrelevant vectors to prevent bloat