Contextual Coherence in Conversational AI: Leveraging a Memory Agent

This study presents a Memory Agent Framework specifically designed for engaging conversational AI while addressing context coherence and memory scalability. The design is based on a Model Context Protocol (MCP) to synchronise many conversational agents over four memory layers inspired by cognitive functions (Short-Term, Episodic, Semantic, and Procedural). The Python-based asynchronous orchestration helps quickly retrieve memories, using FAISS/Pinecone vector storage and a Neo4j knowledge graph to dynamically reason a conversation (like a human would use memories). When the Memory Agent Framework was evaluated in three different modes (baseline, fine-tuned, and embedding-enhanced), it showed a real performance advantage through the Memory Agent Framework compared with baseline evaluation mode. In the Short-Term Memory, each perplexity level decreased from 327.18 to 294.47; in the Episodic Memory, it decreased from 348.49 to 313.64; and in Semantic Memory, it decreased from 344.22 to 309.79. However, the semantic coherence improved by 28.5% in contextual reliability, reaching 0.5499. The fine-tuned models achieve BLEU and ROUGE-L scores above 0.5, indicating improved grammatical correctness and relevance. These results suggest that the Memory-driven paradigm improves multi-turn conversation comprehension, contextual fragmentation, and episodic interaction continuity and understanding. Overall, the Memory Agent architecture supports scalable, context-rich conversation systems that promote coherence and adaptive reasoning in real-world communication.