// the find
caura-ai/caura
Caura (formerly MemClaw) — governed shared memory for AI agent fleets. Multi-agent, multi-tenant, MCP-native. Trust tiers, keystone policies, audit trails, knowledge graph, self-improving retrieval. Apache 2.0.
Caura (renamed from MemClaw) is a self-hosted or managed memory layer for teams running many AI agents at once — Postgres+pgvector for storage, Redis for caching, and an MCP server exposing 12 tools for write/recall/governance. It's explicitly aimed at agent fleets, not solo Claude Code users; a single agent gets none of the differentiating features, only the operational overhead.
The entity resolution logic is genuinely well thought through — canonical-name matching strips leading qualifiers like 'the'/'new'/'old' but only when two or more words remain, so 'new york' doesn't collapse into 'york'. Contradiction handling is a real guarantee, not a heuristic: when a stale memory and its correction both surface in a result set, the correction is ranked above it every time. The Broker's cloud wire protocol is frozen at v1 with an oasdiff breaking-change gate in CI, which is more API discipline than most infra projects this size bother with. Telemetry disclosure is unusually thorough — exact payload shape, four independent opt-out paths, and a boot-log line stating the on/off decision on every start.
The MemClaw→Caura rename is only half done in practice: old env vars, tool names, the daemon binary, GHCR image names, and even the installed skill path (~/.claude/skills/memclaw/) are all still MemClaw, so a new user is tracking two names for every concept. Every write triggers a synchronous LLM call for classification, title, summary, and entity extraction — a real latency/cost tax per memory — and self-hosting without a provider key falls back to dummy embeddings with no semantic search. The comparison table against Mem0/Zep/Letta is written by the project itself from 'a reading of public docs,' so treat those checkmarks as marketing rather than a benchmark. The bigger features — Skill Factory, Interviewer, Broker Fleet — are all opt-in and off by default, and the README describes a lot of lifecycle machinery (six promotion gates, a content scanner, crash-safe windowed ingestion) that's hard to verify is as mature as it reads without actually running it.