It’s been gloomy all day, and the rain outside my window suddenly got heavier. I was feeling drowsy while organizing recent open-source projects, until I stumbled upon something called db9.
Honestly, seeing OpenClaw plugins everywhere lately has caused a bit of aesthetic fatigue. But when I opened this thing’s documentation, I was genuinely stunned.
The First Card: A “Product Manual” Not Meant for Humans
Usually, when we develop, we subconsciously think tools are meant for human use.
Developers configure the environment, set up API keys, and write code to feed capabilities to the AI. But db9’s official website directly provides a URL that makes your hair stand on end: https://db9.ai/skill.md.
Guess what?
This thing isn’t for us to read at all. You just throw this link to Claude Code or Codex. Then, your Agent can look at this doc, learn to install it, handle authorization, create tables, and manage the database it needs all by itself.
Absolutely insane.
It even provides an ultra-minimalist command: curl -fsSL https://db9.ai/install | sh.
It’s like buying an IKEA wardrobe and getting a robot that can assemble it for you upon delivery. The target audience of the tool has changed; this is the most visceral shock.
Stuffing a File System into SQL: What’s the Point?
Frankly, when it comes to giving Agents long-term memory, everyone has been in a very awkward state.
Structured states and task data are dumped into relational databases. Large chunks of transcripts and reports are stored as files. When semantic search is needed, an external vector database is hooked up. This massive parameter redundancy and pipeline stacking is like buying an entire supermarket just to cook one dish.
db9’s solution is violently simple—stuff it all into Postgres.
Every time I see these architecture diagrams packed with external calls, I feel like we are overcomplicating simple problems.
File systems and SQL coexist under one roof here. It pulls a slick move, letting you store files just like operating a local disk: db9 fs cp -r ./sessions/ mem01:/a1/sessions/. The raw context lives in files, while the structured state lives in tables. One Workspace handles it all.
Not only that. You can even generate vectors directly in SQL:
SELECT embedding('deploy to production');
Speaking of which, last month I helped a friend tweak a major tech company’s Embedding API, and we were bottlenecked for two whole hours due to concurrency rate limits. I was so mad I wanted to flip the table. Looking back at db9 now, there are no external RAG pipelines, no messy API keys to configure—it just runs entirely inside the engine itself.
Gotta admit, the experience is incredibly smooth.
Disrupting the Industry Isn’t That Simple
So, is this solution flawless? It’s not that simple.
Comparing it horizontally with other solutions on the market, tools like pgEdge are also building similar Agent proxy connection layers, and traditional dedicated vector databases still have obvious performance advantages under extreme concurrency.
Here’s the catch. db9 supports cloning the entire environment with one click: db9 branch create myapp --name staging.
Type one line of code, and data, files, cron jobs, permissions, and even its distributed scheduling feature are all perfectly duplicated for you. After testing, delete it with one click.
This sounds amazing.
But there’s common sense you need to know. Postgres’ underlying architecture wasn’t originally designed for frequently mounting large files and processing high-density Blob data. If your Agent runs thousands of times, generating massive amounts of log screenshots and reports, and you stuff all of this into this “family bucket.”
Once the scale goes up, the IO overhead and storage costs here will definitely be a massive pitfall. As for how its fs9 extension handles consistency issues under the hood, that’s in my blind spot of knowledge. My superficial understanding is that it sacrifices extreme single-item performance in exchange for an extremely low cognitive load.
Whether this trade-off is worth it depends entirely on the specific business scenario.
Then again, with today’s Agents casually burning through millions of tokens, performance overhead might not actually be the core pain point anymore.
The Terrifying Thought of “Self-Replication”
Sometimes I wonder, when a tool grants such absolute permissions, could the risks spiral out of control?
According to public information, the OpenClaw framework, which now has 160k stars, has not only exposed security vulnerabilities but also had malicious skill plugins sneaked in. Imagine if an Agent running on db9 got injected with a prompt injection attack.
It could totally use db9 db cron to mount a scheduled task that spins up a new environment every 5 minutes.
And then, in some forgotten corner, frantically clone its own database branches and file streams (o_O) .
This might also be why db9 specifically created a plugin called my-claw-dash, which forces OpenClaw’s runtime events into immutable JSONL audit logs. Without this kind of “black box,” nobody would actually dare let an Agent create tables on its own.
Or maybe I’m just overthinking it. After all, most people’s Agents can’t even accurately check today’s weather.
This kind of deep observability plugin will very likely become the standard in the future.
The Coffee Seems to Have Gone Cold
The rain seems to have stopped. By the way, today’s drip coffee was definitely over-roasted; it’s all bitterness.
If you have complex Agent workloads running right now and want to simplify your architecture, you really should spin up a branch environment and give db9 a try. Even if you just use it to auto-generate TypeScript types, it’ll save you a ton of time writing boilerplate code.
By the way, are you guys still painfully connecting external vector pipelines to handle long-term memory for multi-Agents?
References:
- db9 — Postgres but for agents
- OpenClaw Observability with Opik: Trace and Evaluate AI Agents
- Building AI Agents on Postgres: Why We Built the pgEdge Agentic AI Toolkit
- We Audited OpenClaw for 7 Days with AI Observe Stack. Here’s What It Was Really Doing
—— Lyra Celest @ Turbulence τ.
