MongoDB stores flexible documents; PostgreSQL is a relational database with strong SQL and JSON capabilities. AI does not make that distinction disappear. Choose based on the durable application data and access patterns surrounding the model.
Short answer
PostgreSQL is the default for AI SaaS products with users, organizations, subscriptions, permissions, runs, and audit relationships. MongoDB is attractive when the dominant unit is a variable document and the team benefits from its document model and operational ecosystem.
| Workload | Likely starting point | Why |
|---|---|---|
| Users, tenants, billing, permissions | PostgreSQL | Constraints and relational queries |
| Variable event or source documents | MongoDB | Natural document representation |
| Conversations | Either | Depends on query and update patterns |
| Agent runs and tool audit | PostgreSQL often | State transitions and relationships |
| Mixed structured and flexible metadata | PostgreSQL JSON or MongoDB | Benchmark actual queries |
Conversations are not the whole schema
A chat transcript may look like one nested document, but production systems also query messages by user, retention class, incident, model, tool, and cost. Large documents can create contention or awkward updates. In PostgreSQL, rows and JSON columns can combine relational ownership with flexible provider metadata.
Agent state
Persist run status, checkpoints, tool calls, approvals, idempotency keys, and errors explicitly. Do not use an ever-growing conversation blob as a workflow engine. Either database can store state; the important requirements are atomic transitions, concurrency control, indexing, and recovery.
Retrieval metadata
Keep source identity, tenant, permissions, version, and deletion state alongside chunks or references. Vector support is optional and workload-specific. Do not add embeddings to the primary database merely because an extension or index exists; test filtered recall, write volume, index maintenance, and backup behavior.
Consistency and schema evolution
MongoDB schema flexibility moves validation responsibility into applications and migrations; it does not remove schema. PostgreSQL constraints can prevent invalid relationships but require deliberate migrations. For model-generated data, validate before writing regardless of database.
Scale and operations
Both systems can serve large workloads. Compare managed-service limits, failover, backups, point-in-time recovery, connection behavior, geographic requirements, observability, and team expertise. “Web scale” is not a decision criterion.
Decision rule
Start with PostgreSQL when the product is relational and correctness-sensitive. Choose MongoDB when document access patterns are genuinely central and proven. Keep object files outside either primary database and design queues separately. See AI Data Infrastructure and agent memory patterns.