Curltale AI

Governed meaning · Verified knowledge

Memory for your business data.

AI agents can write SQL and answer questions. They cannot know what your words mean or what your documents really say. Curltale AI gives them both, with proof.

View on GitHubSee how it works

Open source (MIT). Self-hosted. Your data stays in your database.

Scattered facts being bound into one connected memory: data-model entities (squares) linked to topics from documents (circles) CustomerOrderInvoice Pricing rulesRefund policyPlan terms ontology knowledge a link that stays in the collection

The problem

AI is fluent. It is not informed.

Without your context an agent guesses, and its guesses look right.

Whose “active”?

Three definitions exist. The agent picks one.

Revenue that doubles

Two tables joined flat. A confident, wrong number.

Last spring's policy

Quoted from a stale PDF. Nobody notices.

What you need

Four things between your data and an AI you can trust.

You bring the data and the models. Curltale AI adds the layer in between.

What you need between your data and trustworthy AIAgents and apps ask at the top. Your databases, documents and people are at the bottom. In between sits the context layer Curltale AI provides: a semantic layer, verified knowledge, governance, and an open agent interface.WHO ASKSpeople and AI agentsTeamsAI agentsAssistantsYour appsThe context layerCurltale AI adds all four on top of your dataSemantic layerBuilt on your databases:definitions, joins, rulesVerified knowledgeWhat your people know,each fact with its proofGovernanceWho may see and changewhat, with an audit trailAgent interfaceOne open endpoint (MCP)for any assistant or toolYOU BRINGyour data stays where it is: read, never copiedYour databasesYour documentsWeb pagesWhat experts know

Where we fit

The layer BI tools and search do not cover.

A semantic layer knows your data but not your documents. Search knows your documents but not your data. Curltale AI connects both, and proves what it says.

Where Curltale AI fits among what companies usually buyA comparison of a BI semantic layer, RAG or vector search, a wiki or knowledge base, and Curltale AI across six capabilities: knowing what data means, reading documents, checking facts, tying documents to data, controlling agent access, and serving agents through MCP. Curltale AI covers all six.A BIsemantic layerRAG orvector searchA wikior knowledge baseCurltale AIKnows what your data meansReads your documentsChecks facts against the sourceTies documents to the data they concernControls what each agent may seeServes any AI agent through MCPbuilt inpartly, or with extra toolsnot its jobTypical capabilities of each category. Individual products vary.Curltale AI builds its own model on your database. It does not replace your data platform.

The ontology

From scattered data to answers you can trust.

Your data is already in tables. What is missing is what it means. Curltale AI writes that down once, with your team, and enforces it for every question.

From scattered data to answers you can trustFive steps: connect a database or a spreadsheet read-only; describe entities, fields and joins, drafted for you; define your words as exact conditions and rules; a person reviews and publishes, every version kept; agents ask and get one checked statement, run read-only.01ConnectDatabases or aspreadsheet, read-only02DescribeEntities, fields andjoins, drafted for you03DefineYour words as exactconditions and rules04ReviewA person approves,every version kept05AskOne checked query,run read-onlyEnforced per columnVersioned and auditedChecked on real dataReady for AI
Meaning

Written down once, by the people who know

What an active customer or an open order is, how tables join, which columns are sensitive. Start from a spreadsheet or read your database, then your team adds the definitions.

Answers

Correct by construction

An agent asks for a plan, not for guesses: the right tables, joins and filters, in one statement, checked against the caller's access and run read-only.

One entity, fully described: meaning, fields, allowed values, sensitivity
One entity, fully described: meaning, fields, allowed values, sensitivity.
The model as a graph: how your business objects connect
The model as a graph: how your business objects connect.

The knowledge base

From scattered documents to knowledge you can check.

What your people know sits in files, web pages and heads. Curltale AI turns it into pages an agent can rely on, and shows the proof behind every fact.

From documents to verified knowledgeFive steps: capture files, web pages and interviews; write a clean page per document; verify it claim by claim against the original; a person reviews what was doubted; publish topics, claims and links for agents.01CaptureFiles, web pagesand interviews02WriteA clean page perdocument03VerifyChecked claim by claimagainst the original04ReviewA person decideswhat was doubted05PublishTopics, claims andlinks, for agentsTraceable to the sourceEvery version keptGovernedReady for AI
Verified

Checked, not just indexed

Each document is rewritten as a clean page, then checked claim by claim against the original by a second model. Anything doubtful waits for a person.

Proof

Every fact keeps its quote

Claims carry their exact quote, the source and the page version. When new evidence contradicts a published fact, the change is held, with both versions side by side.

A topic page: the article with its citations, and beside it every claim with its exact quote and source
A topic page: the article with its citations, and beside it every claim with its exact quote and source.
New evidence that contradicts a published fact is held: the reviewer sees both versions and the quotes
New evidence that contradicts a published fact is held: the reviewer sees both versions and the quotes.

See it

Ask. Get the answer and its proof.

The count comes from the data, with its definition
The count comes from the data, with its definition. The refund rule comes from the documents, and where two sources disagree the assistant says so.

Why it is different

Beyond traditional RAG.

The careful work happens once, with people in the loop. Not at question time, by a model guessing.

Beyond traditional RAGTraditional RAG returns raw document fragments with little business context; Curltale AI returns whole verified topic pages, understands business words through the ontology, and every fact is traceable.Traditional RAGReturns raw document fragmentsLittle business contextAnswers drift between askersHard to trace and governVSCurltale AIReturns whole, verified topic pagesKnows what your words meanGrounded, consistent answersEvery fact traceable to its source

Trust and control

Your infrastructure. Your rules.

Enterprise trust and controlSelf-hosted and open source; access and governance with roles, grants and review; data protection with read-only connections, column sensitivity and an audit log; built on Docker, PostgreSQL with pgvector, six database connectors, any OpenAI-compatible model and MCP.Self-hosted, open source (MIT)runs in your environment, under your controlAccess and governanceroles, collection grants, review firstData protectionread-only, column sensitivity, audit logON YOUR TERMSDocker deploymentPostgreSQL with pgvectorSix database connectorsAny OpenAI-compatible modelMCP for any agentEvery call in the activity logone image, optional worker and storagevectors live beside your dataPostgreSQL, MySQL, MariaDB, ClickHouse…hosted or self-hosted, one per jobone endpoint, read-only toolswho asked, what it cost

In practice

What teams build with it.

Customers

Assistants that stay on message

Answers only from collections you mark public.

Analytics

Self-service answers

“Active” means what the business means.

Compliance

Policy answers with evidence

Every claim carries its quote and source.

Experts

Knowledge that outlasts the person

An interviewer captures it before it walks out.

Try it on us

Ask Curltale about Curltale.

Our own documentation runs on Curltale AI, and it has a public MCP server. Point your agent at it and ask how to install, model and run the product. It answers from the docs, on the product it describes.

claude mcp add --transport http curltale-docs https://mcp.curltale.com/mcp \
  --header "Authorization: Bearer <public-token>"

Read-only, no data, nothing to sign up for. The public token and a ready prompt for your agent are on the docs page.

Connect your agentOr start with your first hour

Get started

Open source. Running in an afternoon.

Two containers, any AI model you choose, and your data stays where it is.

git clone https://github.com/curltale/curltale.git && cd curltale
cp docker-compose.example.yml docker-compose.yml
cp .env.example .env
docker compose up -d
View on GitHubRead the guideReleases

The name

Why Curltale?

The hippocampus is the part of the brain that turns experience into lasting memory. It binds the pieces of what happened into one connected whole, and brings that whole back when you need it. Its name is Greek for seahorse, because its curled shape looks like one. AI agents have nothing like it: without something steady to hold on to, they guess. Curltale is a hippocampus for AI agents. It holds on to your ontology and your verified facts, and turns your data and documents into a tale an agent can tell again, with proof.

  • The shape is real. Hippocampus is Greek for seahorse, named for its curled form. Its other old name, Cornu Ammonis, means “Ammon's horn”, a ram's horn. Both are spirals.
  • It consolidates. The hippocampus turns fleeting experience into durable memory. Curltale turns raw documents into verified pages, each fact with its exact quote.
  • It binds. The hippocampus ties the details of an event together. Curltale ties your data and your documents together, so a policy points at the things it governs.
  • The tail is our own picture. The brain structure is named for its shape. A seahorse also wraps its tail around something steady so the current cannot carry it away, and that is the image we chose for what Curltale does: hold on to verified facts instead of drifting into guesses.

Hold on to the story.