Shared memory for AI fleets that actually governs

What it does — Caura is a memory layer for multi-tenant AI agent fleets. Agents submit plain text through MCP or REST; the system enriches each write into a searchable memory, auto-detecting PII, extracting entities, and classifying it into one of fourteen memory types. Fleets can then recall knowledge across scoped visibility levels—private to an agent, shared within a team, or governed across the entire organization—while tenant isolation is enforced at the database row level.

The interesting bit — The search ranking is deliberately opinionated: it composites pgvector similarity, full-text matching, knowledge-graph expansion up to two hops, and a freshness boost, but if a memory has been superseded, the stale version is allowed to surface yet is always ranked immediately below its replacement. Governance is also baked in from the first write via four agent trust tiers and visibility scopes, so cross-fleet recall is permissioned rather than open-season.

Key highlights —

• MCP-native tools (caura_write, caura_recall) and REST API, with Python and TypeScript clients.
• Hybrid retrieval combining pgvector semantic search, full-text keyword matching, and live knowledge-graph expansion.
• Single-pass LLM enrichment derives memory type, title, summary, importance score, and PII flags from a single content field.
• Production usage at eToro cites 300+ agents, 26,500+ memories, and 23 ms p50 search latency.
• Runs self-hosted via Docker Compose with Postgres and pgvector, or as a managed platform; a local embedder option keeps semantic search fully offline.

Caveats —

• Standalone mode ships with dummy embeddings, so semantic paraphrase search and model-inferred metadata require configuring an external AI provider.
• Auto-registration on first write is treated as a legacy fallback; the docs recommend atomic agent provisioning via POST /admin/agent-keys/provision instead.

Verdict — Evaluate it if you are running multiple agents that need to share operational knowledge under governance. For a single chatbot with no compliance or fleet-growth concerns, it is likely overkill.