Developer moves AI swarm memory from Zep into PostgreSQL to cut costs and dependencies

A developer building a multi-agent simulation gateway called MiroFish faced a recurring cost problem, with Zep projected to consume 30,000–40,000 credits per simulation run. To avoid depending on an external provider's availability, billing, and data isolation, the developer migrated swarm memory — including graph entities, agent episodes, and semantic embeddings — into a self-hosted PostgreSQL 16 database using the pgvector extension. The solution covered three core read patterns: entity retrieval by type, bounded graph traversal, and meaning-based episode search, all within a single transactional store. Tenant isolation was enforced by requiring a tenantId argument on every memory query, with unique constraints preventing duplicate entries per simulation. The developer noted the approach suits bounded graph traversals at current scale, but acknowledged that larger, more complex path analysis may warrant revisiting the architecture later.