Write. Search. Fork. The State Store for AI Agents.

MySQL-compatible · Embedded or Server · Hybrid Vector + Full-text Search · COW Sandbox

⚡ 1,523 QPS streaming write+search (10× Milvus, 3× Elasticsearch)
🌿 FORK/MERGE sandboxes for safe agent exploration
🔍 Vector + full-text + scalar in one SQL query
🐬 Full ACID, MySQL protocol, works with LangChain/LlamaIndex/Dify

English | 中文版 | 日本語

30-Second Try · Quick Start · Why seekdb · Ecosystem · Contributing

_{If you find seekdb useful, consider giving it a star — it helps others discover the project.}

---

⚡ Performance at a Glance

📖 Read the launch blog → · 🔁 Reproduce the benchmark →

⏱️ 30-Second Try

pip install -U pyseekdb   # pyseekdb is the Python SDK for seekdb

📋 View demo.py source →

No servers, no schemas, no embedding setup. Embedded mode runs in-process; switch to server / OceanBase mode with one line. More examples →

✨ Why seekdb for Agents?

🔥 Streaming Write + Concurrent Search, Without the P99 Spike

Agent workloads are continuous write + millisecond-later read. seekdb's async index pipeline (Change Stream) decouples DML from index build, and its two-level HNSW (incremental + snapshot) makes newly-written vectors immediately searchable.

The write path commits and returns without waiting on index construction. The Change Stream pipeline consumes the redo log asynchronously and updates the delta HNSW. Queries hit both delta and snapshot indexes with fine-grained read locks — this is why P99 stays flat under concurrency.

The result: 1,523 QPS with 21.7 ms concurrent P99 — 10.7× the QPS of Milvus, and P99 jitter of just 1.1× when concurrency rises (vs ~10× for ES / Milvus on the same workload).

_{Source: src/share/change_stream/ · src/share/vector_index/}

🌿 Copy-on-Write Sandboxes for Agent Exploration

FORK DATABASE snapshots an entire database in seconds — no data copy. Agents experiment freely (write, query, even break tables); then MERGE TABLE commits the work back, or DROP DATABASE discards it. Kernel-level COW, not application-layer save/restore.

-- Snapshot in seconds, no data copy
FORK DATABASE agent_state TO agent_sandbox_42;

-- Agent reads/writes freely on the sandbox...
USE agent_sandbox_42;
INSERT INTO memory (session_id, embedding, content) VALUES (...);

-- Accept the work back to mainline (strategies: FAIL / THEIRS / OURS)
MERGE TABLE agent_sandbox_42.memory INTO agent_state.memory STRATEGY THEIRS;
-- ...or throw it away:
DROP DATABASE agent_sandbox_42;

_{Source: tools/deploy/mysql_test/test_suite/fork_table/}

🔍 Hybrid Search in a Single SQL

Vector + full-text + scalar filter pushed into one execution plan. No N+1 client-side merging, no glue code to combine results.

SELECT id, title, l2_distance(emb, '[0.12,0.34,...]') AS dist
FROM docs
WHERE MATCH(content) AGAINST('quarterly report')
  AND author_id = 42
  AND created_at > '2026-01-01'
ORDER BY dist APPROXIMATE LIMIT 10;

🐬 MySQL-Compatible, ACID, Embeddable

Built on the proven OceanBase SQL engine. Works as an embedded library, a single-node server, or in the OceanBase distributed cluster. Full ACID, real-time writes, and the entire MySQL ecosystem out of the box.

🎬 Quick Start

Installation

Choose your platform:

curl -X POST https://d0.seekdb.ai/api/v1/instances

pip install -U pyseekdb

docker run -d \
  --name seekdb \
  -p 2881:2881 \
  -p 2886:2886 \
  -v ./data:/var/lib/oceanbase \
  oceanbase/seekdb:latest

# Linux (one-line install, may need sudo)
curl -fsSL https://obportal.s3.ap-southeast-1.amazonaws.com/download-center/opensource/seekdb/seekdb_install.sh | bash

# macOS (Homebrew)
brew tap oceanbase/seekdb
brew install seekdb

📝 More Examples

For the full Python SDK walkthrough — connection modes, embedding functions, metadata filters — see the pyseekdb User Guide.

import pyseekdb

client = pyseekdb.Client(path="./agent_state.db")
memory = client.get_or_create_collection(name="episodic")

for step in agent.run():
    # Persist the observation
    memory.upsert(ids=[step.id], documents=[step.observation])

    # Retrieve relevant context — milliseconds after the write,
    # served by the incremental HNSW (no waiting on a background rebuild)
    relevant = memory.query(query_texts=step.next_query, n_results=5)

    agent.act(relevant)

-- Table with vector column, full-text index, and HNSW vector index
CREATE TABLE articles (
  id        INT PRIMARY KEY,
  title     TEXT,
  content   TEXT,
  embedding VECTOR(384),
  FULLTEXT INDEX idx_fts (content) WITH PARSER ik,
  VECTOR   INDEX idx_vec (embedding) WITH (DISTANCE=l2, TYPE=hnsw, LIB=vsag)
) ORGANIZATION = HEAP;

-- Hybrid search: vector similarity + full-text match in one query
SELECT id, title,
       l2_distance(embedding, '[0.12, 0.34, ...]') AS dist
FROM articles
WHERE MATCH(content) AGAINST('quarterly report')
ORDER BY dist APPROXIMATE
LIMIT 10;

📚 Use Cases

🌟 Ecosystem & Integrations

🌐 Next Steps & Community

• 📖 Read the docs → — Quickstart, API reference, integration guides
• 📝 Launch blog → — The architecture behind 10.7× the QPS of Milvus
• 🐛 Open an issue → — Report bugs, request features
• 🤝 Contribute → — Help build the agent-era state store

🛠️ Development

Build from Source

Before building, please install the required toolchain and dependencies for your operating system. See Install Toolchain for detailed instructions.

# Clone the repository
git clone https://github.com/oceanbase/seekdb.git
cd seekdb
bash build.sh debug --init --make
mkdir -p ~/seekdb/bin
cp build_debug/src/observer/seekdb ~/seekdb/bin
cd ~/seekdb
./bin/seekdb

In this example, the working directory is $HOME/seekdb, please use a fresh directory for testing. Please see the Developer Guide for detailed instructions.

Contributing

We welcome contributions! See our Contributing Guide to get started.

📈 Star History

If seekdb is useful to you, a star helps others find it. ⭐

📄 License

seekdb is built by the OceanBase team — the same database engine running in production at Alipay, Taobao, DiDi, Xiaomi, and more. Fully open-source under the Apache License, Version 2.0.