Engram

This gives Claude a persistent memory layer backed by SQLite with semantic vector search and a knowledge graph. It exposes 17 MCP tools including engram_remember for storage, engram_recall for semantic search, engram_consolidate for LLM-powered pattern extraction, and engram_surface for proactive context pushing. The consolidation runs automatically between sessions to distill episodic memories into durable knowledge and resolve contradictions. It also ships with a REST API if you need to integrate from non-MCP environments. The benchmark claim is notable: 80% recall accuracy on LOCOMO versus 28.8% for flat file approaches, using 776 tokens per query instead of dumping full context. Reach for this when you want memory that persists across projects and improves through automatic consolidation rather than manual curation.