caura-ai/caura-memclaw

What it solves

AI agents often operate in isolation, repeating mistakes or failing to share knowledge with other agents in the same organization. MemClaw addresses the need for a centralized, scalable, and secure memory layer that allows entire fleets of agents to learn from each other, share skills, and maintain a consistent knowledge base.

How it works

MemClaw uses a "write, recall, compound" loop to manage information. When an agent writes plain text, the system uses an LLM to enrich that text with titles, summaries, tags, and importance scores. It employs hybrid search—combining semantic similarity via pgvector, full-text keyword matching, and knowledge graph expansion—to retrieve relevant information. The system also features contradiction detection to resolve conflicting information and uses outcome-based learning to reinforce successful agent actions.

Who it’s for

It is designed for developers building multi-tenant, multi-agent AI systems where governance, security, and cross-agent knowledge sharing are critical, such as enterprise-scale agent fleets.

Highlights

• Fleet-wide Governance: Features tenant isolation, visibility scopes (agent, team, or org), and agent trust tiers to control information access.
• Self-Improving Memory: Automatically detects contradictions, merges duplicate memories into canonical facts, and tunes retrieval based on agent feedback.
• Rich Memory Pipeline: Performs single-pass LLM enrichment and maintains a live knowledge graph of extracted entities and relations.
• Extensive Integrations: Includes a built-in Model Context Protocol (MCP) server and supports multiple LLM providers like OpenAI, Gemini, and Anthropic.