Memory for AI agents.
In a bucket you own.
Agents are only as good as what they remember. Polign builds a durable, searchable memory layer that lives in your own object storage. The serving nodes are stateless, so you can scale them up, down, or to zero at a fraction of the cost, while your data stays under your own control.
What we're solving
Agents put demands on a database that the current generation of vector stores wasn't built for.
Memory availability
Agents' memory lives in your cloud bucket, not on local disks or third-party servers. Accessing it is as simple as spinning up a node, and it serves the same collections from the get go.
Fraction of the cost
Agent memory is mostly cold, yet most stores keep it in expensive always-on memory and bill you around the clock. The cost adds up with per-query charges and the compute that never sleeps.
Your data stays yours
Everything sits in a bucket you own, under your own account and keys. Sensitive memory never has to move into third-party infrastructure.
How we solve it
One product, built for exactly this. The whole loop runs inside your own premises, and the bucket stays the single durable copy.
polign_db
A vector and hybrid search database that runs entirely in your own object storage. Stateless nodes keep your hot memory cached close to the query, so recalls are answered locally inside your own network. The bucket is only touched for writes and the occasional cold read, which is where per-query costs disappear.
Run one node on a single host or a fleet as a cluster, scale them up, down, or to zero as load changes, and pay for the compute you run, not per query.
Each node keeps your hot memory cached, so recalls are answered locally at in-memory speed. Nodes are lightweight: spawn more to scale up, wind down to zero when idle. The memory stays in your bucket.
One binary. Your bucket.
Install the server, point it at object storage you own, and your agents have memory that survives every restart.
# install the server and CLI curl -fsSL https://get.polign.com | sh # run a node against your bucket polign-server -store s3://my-bucket/polign # remember something curl -X PUT localhost:23000/v1/collections/memory/vectors/m1 \ -d '{"values":[0.12,0.83,...],"metadata":{"text":"prefers dark mode"}}' # recall it curl -X POST localhost:23000/v1/collections/memory/query \ -d '{"values":[0.11,0.80,...],"k":5}' {"hits":[{"id":"m1","distance":0.02,"metadata":{"text":"prefers dark mode"}}]}
- Durable before acknowledgedA write is confirmed only once it is in your bucket's log, in one order per collection.
- Read your writesA memory written on one step is readable on the next, even from a node that has never seen it before.
- Search three waysBy meaning, by keyword, or both fused in one query, with metadata filters on top.
- Nodes you can throw awayServing nodes are stateless caches. Add them when traffic spikes, remove them all when it stops.
Connect with us
Start with the free download, no account and no sales call. Self-hosting has no license file and no node limit. When you want Polign on your team, there is one paid option, an annual Production Agreement. See pricing.
Questions, design-partner conversations, commercial terms. connect-us@polign.com
GitHub
Releases, issues, and the open source Python client. github.com/Polign/polign
Live demos
Two agents on two nodes sharing one typed memory in a namespace of your own: polign.com/memory-demo. Semantic search over Wikipedia served from object storage: demo.polign.com.
Ready to give your agents a memory?