What is Barrel?
A database that helps AI applications remember, find and work with information.
AI models can write answers, but a real application also has to remember what happened, look up reliable information, keep track of changes and share data between the different agents that do the work. Barrel is the place where all of that is kept.
Take a support assistant. When a customer writes in, it needs the help articles that answer the question, what this customer asked last week, and a note of what it already tried. Barrel keeps all of that, so the assistant finds it in a moment and picks up where it left off.
About a four minute read. Developers can jump to the technical overview.
AI needs memory
A language model knows what it learned during training. It does not know your customers, your documents, or the decision your team made yesterday, and every conversation starts empty. On its own it cannot remember, cannot look anything up, and cannot keep a record of what it did.
So every serious AI application needs somewhere to put information and find it again: the facts it relies on, what each user told it, the work in progress, and a history of what changed. That place is a database.
Barrel is a database built for applications where AI agents do part of the work. An agent is a program that uses an AI model to carry out a task on its own, such as answering a customer, sorting incoming requests or reviewing a document. Several agents often work on the same information, so they need one shared place to keep it.
What is a database, exactly?
A database is software that stores information so it is not lost, lets several programs read and change it at the same time without stepping on each other, and finds what you ask for quickly. Your bank, your email and most websites rely on one. Barrel is a database designed for applications where AI agents read and write the information.
One place for information
Today an AI application is usually assembled from several separate systems, and the application has to keep them consistent. Barrel keeps everything that belongs together, together.
When a document changes, each system has to be updated, and they drift apart.
Change a document and its search information follows. Copy it elsewhere and its files go with it.
An example: a help article on resetting a locked account has its text, a screenshot that goes with it, and the information that lets an assistant find it when a customer writes “I can’t log in”, even though those words are not in the article. In Barrel these three stay together. Change the text and the search information is updated with it; copy the article to another office and the screenshot goes along.
For the technically curious
Each record holds a JSON document, its attachments and a vector embedding under one id. Barrel can compute the embedding when a document is written, using a model you choose, and searches by vector similarity, by keyword ranking (BM25), or both combined. Every version is kept in a history log within a retention window. Read the record and search.
See it work
A support assistant answers a customer. Pick a question, then ask the follow-up: the second answer is better because Barrel remembered the first conversation.
Northwind support assistant talking with Marie, a customer
1 Marie asks
3 The AI model answers
A separate language model writes the reply from what Barrel found. Barrel does not write text.
4 Barrel remembers
After each conversation the application stores a short note. The next question can find it.
A demonstration in your browser with a small set of articles. Real applications choose their own AI model; Barrel stores, searches and keeps the history.
What you can build
These are kinds of applications teams build with Barrel. Each link explains the problem it solves, how it works in practice, and points to an example you can run.
Why Barrel is different
Once an AI application is used every day, the hard questions are rarely about the AI model. They are about where the data lives, what happens when a connection drops, what changed and who changed it, and who is allowed to see what. This is where Barrel helps.
Your data stays where you decide
Barrel runs inside your application or on your own servers. Nothing has to be sent to an outside service, and data can be encrypted on disk.
It keeps working without a connection
Each copy accepts changes on its own, on a server, a device or in a browser, and copies bring themselves back into agreement later. No change is silently lost.
One place instead of four
Text, files, the information needed to search by meaning, and the history of changes live together. Fewer systems to run, nothing to keep in sync by hand.
Nothing is overwritten without a trace
Past versions are kept for as long as you choose, and each change can record who made it. You can see what an assistant knew at a given moment, or try a change on a copy before applying it.
Access is given, not assumed
Each agent or customer can have its own space with its own key. Access is granted for reading or writing, and taken back at any time.
Open source
Barrel is free under the Apache 2.0 license. You can run it, change it and keep using it without depending on a vendor.
For the technically curious
Barrel is written in Erlang/OTP and stores data in RocksDB. It can be embedded as a library, run as a server with a REST API and an MCP endpoint for AI agents, or kept in a browser with the barrel-lite client. Replication is active-active, with conflicts kept rather than discarded; agents get capability tokens scoped to a space. See replication, the agent layer and deployments.
What Barrel is not
- Not an AI model. It does not write text or make decisions. You use the model you choose, from a cloud provider or running on your own machines.
- Not a chatbot. It is the memory and knowledge behind one.
- Not a complete agent framework. Your application or framework runs the agents; Barrel gives them a reliable place to keep and find information.
Barrel and Enki Bridge
Barrel is open source and free to use. Enki Bridge, the European company that builds it, offers help for teams that want it: pilot projects, development support, and managed hosting of Barrel in the EU. Using Barrel never requires it.