How we built a memory system that scales to tens of thousands of agent interactions. Deterministic writes, confidence gating, and graph traversal keep retrieval precise as the corpus grows.
Vector databases are the default for agent memory. We ripped ours out. Here's why a relational knowledge graph on Postgres outperforms pure vector search for agent recall.
Agent memory stores fill up with duplicates, hallucinations, and malformed extractions. Why this happens, why vector databases make it worse, and the deterministic system that fixes it.