Relational Database

Relational databases are suited for scenarios with fixed structures such as user accounts and order records, while non-relational databases are suited for scenarios with variable data structures such as content management and real-time logging.

Graph databases store relationships as first-class citizens, enabling multi-hop relationship queries to be completed in milliseconds via graph traversal algorithms without the need to perform costly multi-table JOIN operations.

Compared with traditional relational databases, vector databases have higher storage and query overheads, and are usually billed by vector dimension, index scale and number of queries.

For teams just starting to build AI applications, starting from an existing relational database plus a vector extension is the lowest-risk choice; for teams whose retrieval scale has clearly reached the hundred-million level and centers on semantic recall, the performance advantages of dedicated vector databases are worth the extra operational complexity.

Decades of technical debt, fragile code dependencies and rigid relational databases leave legacy systems unable to support real-time data ingestion, vector indexing and low-latency API interactions.

Although relational databases support JOIN, queries over multi-hop paths of variable depth require recursive CTEs, and their performance and expressiveness are inferior to those of graph databases.

Traditional relational databases require multiple JOIN operations to express multi-hop associations, and performance degrades sharply as association depth increases; graph databases store relationships as first-class citizens, so multi-hop query latency is approximately linear in path length.

For multi-hop queries, recursive SQL requires a self-join against the same large table at every hop, intermediate results balloon exponentially with depth, and it comes with the overhead of temporary tables, sorting, and hash indexes.

Traditional relational databases use optimistic locking (version number comparison) or pessimistic locking (row-level locking) to prevent the lost update problem.

A more reasonable approach for an Agent memory system is often a hybrid architecture: structured metadata stored in a relational database, semantic content retrieved with the aid of a vector index, and then rule-based filtering to decide the flow of information back.

Structured filtering performs equality or range queries on existing columns in a relational database (such as contract number, customer ID, date, and status fields), and is essentially a WHERE clause.

B-tree indexes are widely used in relational databases because they support equality comparisons, range conditions, and ordered retrieval, making them suitable for identifiers, timestamps, frequently filtered columns, and many foreign key lookups.

Traditional relational databases are good at exact matching (such as WHERE conditions), but lack the ability to efficiently answer "which stored vectors are closest to the query vector", because this is a geometric problem.

Relational databases remain the mainstream choice for enterprise core business systems, due to transaction support, data consistency guarantees, and mature query optimization mechanisms.

Structured storage (such as relational databases or Redis) is suitable for storing metadata such as failure counts, timestamps, and user types.

文档型数据库倾向于将相关数据内嵌在同一文档中，以牺牲部分一致性换取查询效率，而关系型数据库强调数据规范化