What problem does it solve? Periodic audits of a live SQLite memory bank (memory.db) often produce contradictory findings because samplers, timezones, and concurrent writers distort the data. This Skill provides a disciplined read-only forensic workflow that reconciles every claim against direct SQL before reporting. ## Core Features & Use Cases - Ground-truth reconciliation: Opens the bank read-only (mode=ro), normalizes timezones, and validates sampler statistics against direct COUNT(*) queries in one connection. - Churn and dedup forensics: Distinguishes true re-ingest generations from first-time ingests, and separates multi-process race duplicates from blind-insert duplicates using id contiguity analysis. - Self-correction and growth analysis: Audits TTL/sweep degradation paths, access/rating distributions, and decomposes growth into discrete events instead of extrapolating bursts. - Use Case: An engineer auditing an ai-raccoon memory.db finds 14.2% duplicate rows; the workflow traces them to multiple server processes racing a non-atomic check-then-insert and recommends a UNIQUE index fix. ## Quick Start Audit my memory.db read-only and produce a graded report of churn, duplicates, TTL sweep behavior, and growth projections with SQL evidence for every finding.
