SeekDB (OceanBase)
An open-source (Apache-2.0) database pitched as “the AI-native state store for agents” — “MySQL-compatible, embedded or server, hybrid vector + full-text search, COW sandboxes (FORK/MERGE).” Built by the oceanbase team (whose distributed database powers Alipay/Taobao), SeekDB is the first agent-memory / state-store infrastructure-layer product in this wiki — the persistent substrate beneath the durable-agents / self-improving-agents threads, not another harness.
Why “for agents” — the design points
- COW sandboxes —
FORK/MERGE. Copy-on-write branches let an agent fork the state, explore/try changes in isolation, then merge back or discard — i.e. branchable, rollback-able state. This is the storage-layer analog of the exploration/reversibility discipline the spoke tracks in agent-guardrails and the fork-and-try patterns of agent-orchestration: cheap-to-undo experimentation pushed down into the database itself. - Hybrid search in one SQL query — vector similarity + full-text matching + scalar filters, combined. The retrieval substrate an agent’s memory/RAG layer needs without bolting a vector DB onto a relational one (cf. research-wiki’s hybrid-retrieval-for-RAG thread, via the gbrain bridge).
- Real-time write→read. An async index pipeline decouples writes from indexing and a two-level HNSW vector index makes newly-written vectors immediately searchable — built for the agent loop’s “write a memory now, retrieve it milliseconds later” access pattern. Full ACID.
- Deploys embedded or server — a library inside the app, a single-node server, or a distributed cluster node; MySQL wire-protocol compatible, so it drops in under agent frameworks LangChain / LlamaIndex / Dify with no new driver.
Claimed performance
“1,523 QPS streaming write+search (10× Milvus, 3× Elasticsearch)” with stable P99 under concurrent load. Caveat: vendor benchmark from the repo, not independently reproduced — the spoke’s standing “READMEs/vendor numbers, no neutral benchmark” caveat applies (see synthesis open questions).
Where it sits
The spoke’s memory thread has so far lived inside harnesses — openhuman‘s SQLite
“Memory Tree”, durable-agents’ workflow-state separation, self-improving-agents’
accumulated memory, gbrain (cross-wiki) as a personal KB used as agent memory. SeekDB
is the first time that layer shows up as a standalone database product purpose-built
for it — the state/memory substrate the harnesses sit on top of. The FORK/MERGE
primitive is the distinctive bet: treat agent state like a version-controlled branch,
not a single mutable blob.
Related
agent-memory · oceanbase · durable-agents · self-improving-agents · openhuman · agent-guardrails · gbrain · agentic-coding-harness