Developer builds AI agent with three separate memory types, boosts SQL accuracy

A software developer published findings on building an AI agent modeled on three distinct human memory systems: semantic, episodic, and procedural. The experiment used a language model tasked with writing SQL queries against the unfamiliar Northwind database, where a strong model succeeded only about 25% of the time without memory support. The developer used Cognee, a knowledge graph tool, to handle semantic memory (schema facts) and episodic memory (past successful queries), while Synapse-DB managed procedural memory through a reinforcement-and-decay mechanism inspired by Hebbian learning. A key finding was that a single vector database cannot distinguish between these three types of not-knowing, leading to recall failures that are difficult to diagnose. The system also incorporated a forgetting step, allowing poorly performing memory buckets to be pruned mid-run to prevent misleading past episodes from corrupting future responses.