SECONDBRAIN V2: MEMFORGE

MemForge is a production-grade personal knowledge graph that runs 10 automated phases every night — blending the precision of SQL scoring with the semantic power of LLM-driven sleep cycles. It self-organizes, self-prunes, and self-revises without manual curation. The system behind over 2,000 notes that just work.

The core architecture: A dual-tier system: Obsidian markdown files as the cold, durable storage layer, and a 60-table SQLite database as the hot scoring/query layer. The two stay in sync through SHA256-based change detection — only modified files get reprocessed.

The scoring engine (6 phases): - Phase 0 — Weight adaptation — Learns from retrieval outcomes, adjusts scoring weights based on what actually gets used - Phase 1 — Scoring — Importance, confidence, graduation, surprise, staleness, and retrieval success tracked per note - Phase 2 — Triage — Graduate promising notes, archive cold ones, detect staleness - Phase 2.5 — Conflict resolution — Detects contradictory note pairs and resolves them via LLM - Phase 3 — Revision — LLM-driven content revision of high-priority notes - Phase 4 — Graph extraction — Entity extraction, relationship mapping, wikilink graph maintenance

The sleep cycle (4 phases): - Phase 5 — Reflection — High-level insight generation, contradiction analysis - Phase 5.5 — Schema detection — Identifies repeated entity pattern schemas across notes - Phase 6 — Cold purge — Hard-deletes archived entries past retention threshold - Audit chain — SHA-256 hash chain for every mutation, fully auditable

Pillar classification: 4-pillar system (Magno/tech/style/tirol) with keyword matching as primary and Ollama semantic embeddings as fallback — notes auto-classify into domain buckets with confidence scoring.

Operational stats: - 2,034 notes in the vault, 60+ scoring tables, 129/130 test suite passing - Runs nightly via LaunchAgent at 3 AM — fully automated, zero manual curation - Ollama nomic-embed-text (768d) for semantic search, FTS5 BM25 for keyword search -Retrieval latency: milliseconds across 2,000+ notes

This is the memory engine powering every conversation — notes are instantly queryable, automatically maintained, and semantically coherent.