# caura [Health: Active]

**Category:** 🧠 Knowledge & Memory  
**Repository:** https://github.com/caura-ai/caura  
**GitHub Stars:** 505  
**Views:** 0  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/caura

## Description
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.

## Claude Desktop Quick Installation
No confirmed install details yet, so no config is published. Follow the setup steps in the repository above.

## Documentation

## What caura does

The caura MCP server provides a shared memory layer for AI agents that need to retain and reuse facts, instructions, and operational lessons. It supports both single-agent and fleet deployments, with tenant, agent, and fleet boundaries available from the first write.

Agents submit plain-text content, and Caura turns it into searchable memory. A second agent can recall knowledge written by another agent when the memory's visibility and trust rules allow it. This supports workflows where one agent records a lesson and another uses it during an incident, deployment, or support task.

Caura belongs in the 🧠 Knowledge & Memory category. Its documented capabilities include trust tiers, keystone policies, audit trails, a knowledge graph, and self-improving retrieval. Memory visibility can be limited to an individual agent, shared with a team fleet, or made available across fleets under governed cross-fleet rules.

## How it works

The core interaction is a write-and-recall loop. A client calls `caura_write` with an agent, fleet, visibility, and content. Another client calls `caura_recall` with an agent, one or more fleet IDs, and a query. Recall results identify the author of shared memories, helping teams trace where a piece of knowledge originated.

A local standalone deployment can operate without an API key. In that mode, Caura stores memories without embeddings and uses keyword search. Adding an AI provider enables semantic search, so queries can match related wording rather than repeating terms from the stored memory. With a provider configured, fields such as titles, summaries, weights, and tags can be inferred by a model; keyless mode uses a deterministic local heuristic.

## Setup and configuration

For local evaluation, clone the repository, copy `.env.example` to `.env`, set `IS_STANDALONE=true`, and start the Docker Compose stack. The stack includes PostgreSQL with pgvector, Redis, the storage service, the REST/MCP API, and a lifecycle scheduler. The documented local endpoint is `http://localhost:8000`.

The managed service is available at `https://caura.ai/mcp`. Configure an MCP client with that URL and an `X-API-Key` header. A tenant-scoped dashboard key requires an explicit `agent_id` on each MCP call; production fleets should use an agent-scoped credential for each agent.

## Tools and capabilities

The caura MCP server exposes the documented `caura_write` and `caura_recall` tools. Together, they support:

- Recording plain-text facts, rules, and operational lessons.
- Recalling memories by keyword or semantic similarity when an embedding provider is configured.
- Scoping memory to an agent, team fleet, or governed organization-level access.
- Sharing knowledge across agents while retaining author information.
- Applying trust and governance controls to shared memory.

The repository also describes REST endpoints for creating memories and searching them, including `/api/v1/memories` and `/api/v1/search`.

## Limitations and notes

Keyless standalone mode is intentionally limited to keyword retrieval and does not provide embeddings. Semantic paraphrase matching requires a configured AI provider. The README notes that public memory benchmarks generally measure a single agent and single user, so fleet governance and workload-specific latency should be validated in the target environment.

Caura was formerly named MemClaw. Existing `memclaw_*` tool calls and supported `MEMCLAW_*` environment aliases remain available, but new configuration should use Caura names. The project is licensed under Apache 2.0.

_Full upstream README: https://allmcps.com/mcp/caura/readme_