Hey everyone 👋 I'm Amir, one of the founders at Papr.

After talking to hundreds of teams trying to ship AI agents, we kept hearing the same painful truth:

Agents hallucinate because they can’t recall the right context.

Conversation history lives in one place, documents in another, structured data in databases… none of it connected. Retrieval works in demos, but in production it returns fragments, not understanding. It's inaccurate and slow.

So we rebuilt the memory layer from scratch. Think of it as a new type of database for AI that connects context across sources, predicts what users want then powers AI agents to surface context and insights before users ask. It's a simple API call to get started over a weekend.

Building memory that gets faster and more accurate with scale

Most memory systems rely on vector search alone. Great for similar text, terrible for connected context. Some add a knowledge graph to capture conversation history - working, episodic, long-term memory. It breaks as soon as you add more data sources.

So we built something new: the Predictive Memory Graph — a layer that maps real relationships across all your data.

A line of code → ties to a support ticket → ties to a convo with an AI coding agent -> ties to a Slack thread → ties to a design decision months ago. Your knowledge becomes one connected story.

Then we went further: Papr anticipates what users will ask next and pre-caches the right context (available in preview today in the Python SDK). When the prediction hits, retrieval drops under 100ms and accuracy jumps.

This is how Papr ranks #1 on Stanford’s STaRK benchmark with 91%+ accuracy — and why it gets better as your memory grows, not worse like we have seen with context rot.

Not just retrieval - actual insights queryable with natural language or GraphQL

Once your data is connected, AI agents can shift from surfacing text to surfacing insights. Papr can answer things like "what themes are emerging in support tickets", "what products should we recommend for this customer based on others like them" or "if the cost rises, how will margin shift next month".

And because Papr is queryable via natural language or GraphQL, teams use the same memory layer for conversational agents, analytics dashboards, insight panels, and recommendations.

This is retrieval → understanding → insight.

Private by design, open by default

Papr includes ACLs, namespace boundaries, and strict permission management from day one. AI agents will only access what you want it to see, and will respect the user's permissions so data doesn't leak across users in the same namespace.

You can run it fully open-source and self-hosted, or use our managed cloud with predictive memory built in — same API, full control.

Thanks for checking out Papr

We built Papr because we believe AI shouldn’t guess — it should be safe, trustworthy and have a heart that can build relationships. It should understand your context, connect the dots, and deliver the right insight before you ask.

We’ll be here all day answering questions about predictive memory, vector + graph database, GraphQL, insights, and building AI agents that actually work in the real world.

We're excited to see the next generation AI apps you will build with Papr 🚀

Really amazing. Can your servers handle huge data retrieval using the GraphQL endpoint?

Also, can I create separate projects/namespaces for different customers?

Where is the source code?