Show HN: A lightweight, stateless database for agent memory

aanuptalwalkar 1 day ago 8 commentsRead Article on polign.com

ZH version is available. Content is displayed in original English for accuracy.

I've been working on Polign and built a small prototype around something I've been thinking about with agent memory. I have built a lightweight/stateless vector db + BM25 search which works really well with typed facts and structured queries.

It uses your own S3, or GCS bucket as primary storage, and restarting a node is fairly quick.

Demo + writeup: https://polign.com/blog-edge-agent-memory

Live search demo: https://demo.polign.com

Docs: https://polign.com

⚡ Community Insights

Discussion Sentiment

100% Positive

Analyzed from 471 words in the discussion.

Discussion (8 Comments)Read Original on HackerNews

Supermancho•about 3 hours ago

There are countless agent memory stores today. Closed source is a non-starter.

anuptalwalkar•about 1 hour ago

Thanks, is the concern being non-starter on the trust side? the binary makes zero outbound calls. It talks to your bucket and nothing else, and that's verifiable with egress monitoring. I'm putting up a security page that spells it out.

Open-sourcing parts of it isn't off the table. I kept the core closed for now to understand the interest in it.

nater5000•about 2 hours ago

So what's stopping me from using one of the million other versions of this that already exist? Or, even better, pointing Fable to your site and having it just develop its own version of this that I completely own?

I'm throwing in a bit of stank in those questions because I don't see anything here that's particularly unique, and keeping this closed and costly is just... absurd. I've vibe coded the same exact thing multiple times. It works, but it's only as useful as the philosophy of the person using it allows it to be.

So why would someone choose your solution? A vector database hosted in S3 is effectively trivial to set up at this point in time. What are you offering that I can't get from prompting a powerful enough model? If the answer to that question isn't the first thing people read when they open this post, then you can just consider all of this worthless and save yourself the hassle.

anuptalwalkar•about 2 hours ago

When I started, I was using what was available out in the world. The vector dbs I used and even the first version I built weren't really optimized to use RAM and resources conservatively. One of the aspects of Polign is it's optimized towards cold queries and only warms the cache when needed, all while keeping the recall accuracy in acceptable range. I accepted slower first-query latency on cold data as the tradeoff. For example, the demo serves 12.5 million passages from S3 with the server idling around 37 MiB RSS (I realized there is a 20 MB gcp sdk in there).

Hosted db have query costs that grow with the corpus, and self hosted have operational overheads. They do add up over time and I wanted cost flat, which took a while to get it right.

If a vibe-coded version covers your use case, you don't need this. I have added my own experience working and optimizing the storage into this, and Im certain it holds up well for a growing corpus and with simplicity to use it across multiple readers.

ajabhish•about 3 hours ago

This looks promising! Curious why you named it `Polign`?

anuptalwalkar•about 3 hours ago

It's wordplay. I wanted it to be lightweight like pollen, but carrying all the information you need to build on top of it. The 'align' part nods to the core vector search aligning with your queries.

unified101•about 18 hours ago

Closed source - for those who maybe miss this. litestream with sqlite vector based design would be more apt for someone.

anuptalwalkar•about 16 hours ago

Thanks for taking a look. Yes, I kept it closed source for now. I'll mention that on the page.

Litestream with SQLite would solve some of these use cases, and backing up to S3 or GCS works well.

Here I was focused on more dynamic workloads. For example, larger indexes and corpus that can be accessed by multiple readers. I didn't want to invest in local storage, so the data is made to stay cold until it's actually accessed into warm cache.

Also, one node can serve all of that in one collection while serving typed agent memory in another.

Perhaps I missed something, but storage, memory, and operational overhead is something I wanted to worry less about.