Memory Guardian released as open-source memory governance layer for AI agents, with conflict detection and explainable retrieval.

Memory Guardian is an active governance layer for AI agent memory that controls what gets stored, retrieved, how memories age, and detects contradictions—with full explainability at every step. It addresses problems like duplicate facts accumulating, conflicting preferences coexisting without resolution, stale memories outranking fresh ones, and opaque retrieval.

Core features include intelligent ingestion (deduplication and type normalization), task-aware retrieval combining semantic similarity, importance, recency (30-day half-life), and frequency signals, and conflict detection via three providers—heuristic polarity analysis (default), NLI using microsoft/deberta-v3-small-mnli, or LLM-based detection. Every retrieved memory includes score_breakdown, decision_trace, and why_selected reasoning.

Designed for regulated environments (wealth management, KYC/AML, claims, credit underwriting, trade surveillance) where memory contradictions can trigger compliance failures or mis-sold products. Memory Guardian persists conflicts as first-class objects, pins formal risk decisions permanently, and decays soft signals automatically—creating an audit trail for review.

Available as open-source software via Docker Compose or local install (requires Python 3.12, PostgreSQL with pgvector). Strict multi-tenant isolation enforced at API layer via X-Tenant-Id header; optional NLI conflict detection requires ~500MB DeBERTa model download.