MemClaw

MemClaw is the governed shared memory layer for multi-agent AI fleets. Instead of each agent operating in isolation, agents write what they learn, recall what the fleet knows, and get smarter with every interaction. Key Features Cross-agent, cross-fleet recall so discoveries compound instead of dying in silos — without leaking what they shouldn't Keystones — mandatory governance rules that every agent obeys at session start, overriding conflicting instructions Contradiction detection — conflicting memories caught via RDF triples + LLM analysis. Knowledge stays clean automatically Per-agent retrieval tuning — each agent optimises its own search profile. Every query gets smarter LLM Crystallizer — deduplicates and merges near-duplicate memories into atomic facts with full provenance 8-status memory lifecycle — memories transition automatically, no manual cleanup What Makes It Different Not a vector DB wrapper — adds fleet orchestration, governance, audit trails, and self-learning on top of pgvector SOC 2 compliant, Apache 2.0 open source MCP-native — works with Claude Desktop, Claude Code, Cursor, Windsurf in 30 seconds Proof it works In production at eToro (NASDAQ: ETOR) — 300+ agents, 26,500+ memories, 23ms p50 search memclaw 77.6% LoCoMo accuracy, 96.6–98.2% token savings vs full context, 11,500 downloads in week one Who it's for Engineering teams running multi-agent AI fleets Enterprises needing governed, auditable agent memory Developers building on Claude, Cursor, or OpenClaw