caura-ai/caura

What it solves

Caura provides a shared, governed memory layer for fleets of AI agents. It prevents agents from repeating mistakes by allowing them to store learned information and share it across a team or organization, ensuring that knowledge gained by one agent is available to others under strict governance and visibility controls.

How it works

Agents write plain text to Caura, which the system then enriches using a single-pass LLM process to classify memory types, generate summaries, and flag PII. It uses a hybrid search approach combining pgvector semantic similarity, full-text keyword matching, and a knowledge graph to retrieve relevant information. The system also features a "Karpathy Loop" for outcome-based learning, where agents report success or failure after acting on memories to reinforce what works and generate preventive rules.

Who it’s for

It is designed for teams deploying multi-tenant, multi-agent AI fleets—such as companies with dozens or thousands of agents—where auditable memory, cross-agent knowledge sharing, and governance are critical for production deployment.

Highlights

• Governed Memory: Implements tenant isolation, visibility scopes (agent, team, or organization), and agent trust tiers to control access.
• Self-Improving: Features contradiction detection that supersedes stale data and "crystallization" to merge duplicate memories into atomic facts.
• Knowledge Graph: Automatically extracts entities and relations from writes to enhance retrieval via graph-boosted search.
• MCP Native: Built-in support for the Model Context Protocol, allowing easy integration with clients like Claude Desktop, Cursor, and Windsurf.
• High Performance: Optimized for low latency (23ms p50 search) and high token efficiency to reduce costs as agent fleets scale.