caura

Caura (formerly MemClaw) is an Apache 2.0-licensed memory system designed for multi-tenant, multi-agent AI fleets. Agents write plain text, and Caura transforms it into searchable, governed, self-improving memory through a write-recall-compound loop. The project emphasizes fleet-scale governance: tenant isolation, visibility scopes (agent, team, org), four agent trust tiers, full audit logging, and PII detection on every write. The memory pipeline includes single-pass LLM enrichment that classifies content into 14 memory types, hybrid search combining pgvector semantic similarity with full-text keyword matching and knowledge graph expansion, a live knowledge graph with entity resolution, and contradiction detection with automatic supersession. Self-improving features include outcome-based learning (the Karpathy Loop), crystallization of near-duplicate memories, and per-agent retrieval tuning. Caura integrates via a built-in MCP server at /mcp (Streamable HTTP), REST API, Python and TypeScript clients, an OpenClaw plugin, and the Rail SDK for wrapping agent turns with memory recall and storage. Deployment options include a managed platform at caura.ai, self-hosted Docker Compose (PostgreSQL + pgvector, Redis, storage service, API), and standalone local mode requiring no API key. The README cites production use at eToro with 300+ agents, 26,500+ memories, and 23 ms p50 search latency. Benchmarks report 77.6% accuracy on LoCoMo and 92.2% on LongMemEval, with token savings of 96.6% and 79.2% respectively.