Corrective memory for AI agents
to make informed decisions.

A correction, step by step

A chat agent approves a refund exception, a voice agent reads it, and a lead revokes it. Every question below has one answer, set by the rule for that fact.

Sep 12  remember  customer-4812 refund_exception = approved   # chat agent
Sep 20  recall    customer-4812 refund_exception              # voice agent sees approved
Oct 2   remember  customer-4812 refund_exception = revoked    # replaces approved

recall now            -> revoked, replaced approved from Sep 12
recall as of Sep 20  -> approved, with the chat message it came from
history              -> approved on Sep 12, revoked on Oct 2

Try it in a minute

One pip install brings the client and the database. No model or embedding API key needed. The quick start walks through history, as-of reads, and connecting your agent.

# pip install polign-recall
from polign_recall import Client

with Client(local_dir="./recall-data") as memory:
    memory.remember("user", "prefers_editor", "vim")
    memory.remember("user", "prefers_editor", "neovim")
    memory.recall("user", "prefers_editor")[0].value   # neovim
    memory.history("user", "prefers_editor")            # vim, then neovim

Runs on the store you already have

Recall is the memory. The vector store underneath is yours to choose, and the records are the same on every one. polign_db is the backend we built for it and the only one with every capability today; a Qdrant you already run works for the memory tools. The table says exactly what each one gives you.

Postgres with pgvector is next, with leases, so resume will work there too. Start on whichever store you have; the records do not change shape when you move.

Resume after a crash

An agent can record its working state and turns to Recall as it works. When its pod dies or its image is upgraded, the next process picks up from those records instead of a snapshot. Read how it works, or use it with LangGraph.

How it is built

Every memory is an event in an append-only log, and every answer is a fold over that log at the instant you ask about. The registry holds the correction rules, the backend holds the records, and recall serve puts keys and namespaces in front of both. Read the architecture, with diagrams.

Become a design partner

Give us your agent and we will wire Polign Recall into it.

• You bring an agent that is in or near production, and a real memory problem
• We do the integration work, architecture help, a Recall deployment in your cloud on polign_db or the store you already run, and migration from your current memory store
• The commitment is one engineer on your side who owns it, and one 30-minute engineering call a week for the first 4 to 6 weeks
• Done means success criteria we write down together before day one, for one production-like agent writing memory into Polign
• Your infrastructure stays in your account and on your bill
• No charge for 60 days, with written success criteria agreed up front. Meet them and it converts to $12k for the next 12 months. See pricing

What ships today

• Go library with no dependencies outside the standard library
• Python client on PyPI as polign-recall
• MCP server through the recall command, read-only or read-write
• recall serve: one server for many agents over HTTP and MCP, with keys bound to a namespace, and Kubernetes workload identity so a pod needs no key at all
• Claude Code plugin for memory across sessions
• LangGraph package for agents that resume after a crash
• Pydantic AI package for agents that remember their users across runs
• LiveKit Agents package for voice agents that remember the caller
• Vapi adapter for caller memory through authenticated webhooks
• Qdrant backend for teams that already run Qdrant, without agent leases or keyword search
• Starter predicates for coding and personal assistants; new ones are defined by first use, or keep a fixed list
• As-of reads that answer what was believed at an earlier time
• Audit bundles with an offline verifier
• Word-overlap search built in; model embeddings optional
• Plain text in, sentences out, optional: with a model configured, Recall works out the facts and decides how each is filed. Without one, your agent proposes facts and Recall validates them
• recall explain shows why each answer is held: the text it came from, the rule that decides it, and its history
• Evidence links: a fact proposed from text keeps the exact words it came from
• Corrections across predicates: the registry pairs predicates that contradict each other (likes and dislikes, owns and gave away), so a later statement under one ends the same value under the other. An agent can name anything else a statement ends. Conflicts that no rule settles are shown, never guessed
• Forget is a withdrawal that keeps the history readable; delete the records in the store to erase them
• Runs where you run it: your laptop, your cluster, your bucket