A database that lets AI agents fork reality

What it does

seekdb is a MySQL-compatible state store built for agent workloads: continuous memory writes, millisecond-later retrieval, and safe experimentation. It runs embedded, standalone, or inside OceanBase clusters, and speaks the full MySQL protocol so LangChain, LlamaIndex, and Dify plug in without wrappers.

The interesting bit

The FORK DATABASE / MERGE TABLE / DROP DATABASE flow is the standout. It snapshots an entire database via kernel-level copy-on-write in seconds — no data duplication — lets an agent read, write, even break tables freely, then commit back with conflict strategies (FAIL, THEIRS, OURS) or discard. That’s branching logic at the storage layer, not application code. The async index pipeline (Change Stream) decouples writes from HNSW construction so vectors are searchable immediately without P99 spikes.

Key highlights

• Hybrid vector + full-text + scalar search in a single SQL execution plan, no client-side merging
• Two-level HNSW (incremental delta + snapshot) with fine-grained read locks for concurrent write+search
• Benchmark claims 1,523 QPS with 21.7 ms P99, 10.7× Milvus and 3.2× Elasticsearch on streaming write+search workload
• Full ACID, embedded or server mode, built on proven OceanBase SQL engine

Caveats

• Benchmark numbers are self-reported with a linked reproduction repo; independent verification isn’t shown
• COW sandboxing is kernel-dependent; behavior and performance likely vary across filesystems and platforms
• Project is young (2,716 stars); long-term stability relative to parent OceanBase codebase is unproven

Verdict

Worth a look if you’re building agents that need durable memory with safe exploration loops, or if you’re tired of gluing vector DBs to relational stores. Skip it if you just need a vanilla vector index — the MySQL compatibility and COW semantics are the whole point, and overkill for simple RAG pipelines.