Corrective memory
for AI agents.
Your agents search for context that looks semantically close and hope it is still true. With Polign Recall, every recall returns the current facts and the corrections behind them, so the model answers from what is known.
warfarin_dose = 2.5 mgcurrent
What we're solving
When an agent fails, hallucinates, or makes assumptions, it's rarely because of the model. They fail because the memory feeding it is stale, contradictory, or wiped out after a crash. Today's models reason remarkably well when they have the right details. The failure happens when the model is given incomplete or outdated information.
Most of today's agents run a similarity search and pass along whatever sounds close to what was asked. The old value next to the new one, a promise next to its revocation, and nothing about the steps taken before the last restart. The model is left to approximate from half-known facts. People don't make decisions that way, and agents shouldn't either.
Conflicting statements
A customer changes plans, a lead revokes an exception, a user switches tools. Each kind of fact declares whether it holds one value or many, so a new value replaces the old one by rule, not by the model's judgment. Fix it once and every agent sees the fix.
Know what was true, and when
A recall is a state of the memory at any point in time. Nothing is erased. A correction adds a statement and forgetting adds a withdrawal, so you can ask what is believed now, what was believed on any earlier date, and which words a fact came from. When a decision is disputed, you can show what the agent knew at the time.
Crash/failure recovery
An agent writes down its goal, plan, and progress as it works. When it crashes, gets rescheduled, or is upgraded, the next process resumes from those records instead of starting blank. No snapshots, and the memory stays in your own bucket.
Who it's for
Teams whose agents talk to the same people again and again, where the facts change along the way and a wrong answer has a cost.
- Customer support: a promise made in chat holds on the phone, and a revoked one stops holding everywhere
- Voice agents: callers are remembered and never have to repeat themselves
- Sales and account teams: plans, contacts, and renewal dates stay current across every conversation
- Coding assistants: your preferences and project conventions carry from one session to the next
- Long-running agents: work survives a crash or a redeploy instead of starting over
- Regulated industries: you can show what the agent knew when it acted, and the data never leaves your account
How we solve it
Polign Recall is the memory your agents use. It runs on polign_db, our engine, which keeps every record in your own object storage and treats its servers as replaceable, so you can scale them up, down, or to zero. The data stays in your account the whole time.
Polign Recall
The memory layer agents share on top of polign_db. Every fact is typed, so a correction follows a declared rule instead of a model's judgment: a new value for a single-valued fact replaces the old one, and a multi-valued fact collects them. Nothing is deleted, so the earlier value stays queryable and you can ask what was believed at any earlier moment. Agents in different processes and on different machines read and write the same memory, in a bucket you own. Open source under Apache 2.0.
Recommended store: polign_db
The backend Recall was built on, and the one with every capability. It runs on your own compute and keeps its durable state in your bucket, so a node can be replaced or scaled to zero without losing anything. Layered caches keep repeat queries fast. A write is acknowledged only after it is durable in your bucket, and the next vector query on any node sees it. With encryption turned on, each node encrypts data before it leaves the node, using keys you hold.
From the blog
Connect with us
Start with pip install polign-recall, or try the live demo. Self-hosting is
free. See pricing for production support and the design partner program.
Questions, design-partner conversations, commercial terms. connect-us@polign.com
GitHub
Recall's source, examples, and issues, open source under Apache 2.0. github.com/Polign/recall
Live demo
- polign.com/memory-demo Two agents on two nodes sharing one typed memory in a namespace of your own.
Ready to give your agents a memory?