Current AI systems suffer from a fundamental limitation: they lack persistent, structured memory that mirrors human cognitive processes. Vector databases provide similarity search but discard temporal context, emotional weight, and relational structure. This paper presents YantrikDB, an embedded cognitive memory engine that unifies five index types---vector (HNSW), knowledge graph, temporal, decay heap, and key-value---within a single embedded database. YantrikDB introduces three novel contributions: (1) a multi-signal retrieval scoring function with relevance-gated importance amplification that prevents high-importance memories from dominating low-relevance queries; (2) a contradiction-aware memory lifecycle with CRDT-based replication, where conflicts are first-class data structures triaged by severity; and (3) a proactive cognition loop that autonomously detects patterns, surfaces decaying memories, and generates behavioral triggers without external prompting. I evaluate YantrikDB on a synthetic benchmark of 40 golden queries across 12 retrieval categories, demonstrating that multi-signal scoring with graph expansion outperforms pure vector similarity baselines. The engine is implemented in 16,000 lines of Rust with Python bindings and operates as an embedded library requiring no external services. YantrikDB is designed for AI companions, autonomous agents, and any system that must remember, learn, and act on accumulated experience.