Skip to main content
Connect

Use RushDB with AI Agents

RushDB gives agents a structured data layer, a traversable graph, and persistent memory. Choose the shortest setup that matches your agent.

Fastest Setup: Give the Agent the Guide​

For an MCP-capable agent, send:

Fetch https://rushdb.com/agent-setup and follow the instructions exactly.

The machine-readable setup guide walks the agent through MCP connection, schema discovery, memory labels, nested record linking, recall patterns, and semantic search.

Coding Assistant: MCP + Skills + Memory​

For Claude Desktop, Cursor, VS Code Copilot, OpenClaw, or another local coding agent:

  1. Connect the local MCP server — to RushDB Cloud with an API key, or to your own engine in direct mode.
  2. Install RushDB Agent Skills.
  3. Start a new agent session.
  4. Send:
Set up RushDB as my persistent memory layer.

The agent calls getSchemaMarkdown, creates a SESSION, and validates recall. For the full bootstrap prompt and recommended labels, follow the Agent Memory Quickstart.

Custom Agent Harness​

For an application-controlled agent, choose the interface based on where orchestration lives:

InterfaceUse it when
MCP toolsYour framework already supports MCP tool discovery and invocation
TypeScript SDKYour harness runs in Node.js, the browser, or an edge-compatible runtime
Python SDKYour harness runs in Python
REST APIYou need language-neutral HTTP integration with RushDB Cloud
Native protocolYou run your own engine and need language-neutral HTTP integration

Connect with an SDK →

Native Memory Providers​

MCP exposes model-directed tools. Native providers additionally recall before inference, persist completed turns, participate in session/compaction lifecycle, and fail open when remote memory is unavailable.

RuntimeNative integration
OpenClawAdditive RushDB memory plugin — preserves local Markdown/SQLite memory
Hermes AgentRushDB MemoryProvider — automatic prefetch and non-blocking durable sync

Both build on @rushdb/agent-memory-contract, the provider-neutral TypeScript boundary for agent memory: the AgentMemoryEvent v1 model, deterministic ids, scope filters, idempotent writes and semantic recall. Use it to build a native integration for another runtime; see the Agent Memory Event Contract.

Recall is a vector search. On RushDB Cloud, memories are embedded for you. On a self-hosted engine, the integrations need an embedding model of your own: the native providers take an OpenAI-compatible endpoint, and the contract takes an embedder:

import { RushDBAgentMemory } from '@rushdb/agent-memory-contract'

const memory = new RushDBAgentMemory({
  engineUrl: 'http://127.0.0.1:3000',
  engineToken: process.env.RUSHDB_ENGINE_TOKEN,
  embedder: {
    dimensions: 1536,
    embed: async (texts) =>
      (await openai.embeddings.create({ model: 'text-embedding-3-small', input: texts })).data.map((d) => d.embedding)
  }
})

await memory.ensureIndexes() // creates external indexes of the embedder's dimensions

At the start of each agent session:

  1. Call getSchemaMarkdown to discover the live schema.
  2. Recall the latest SESSION and relevant DECISION, TASK, or PREFERENCE records.
  3. Create a new SESSION.

Before the session ends:

  1. Store decisions, tasks, observations, and artifacts.
  2. Link them to the session through nested JSON or relationships.
  3. Confirm the stored summary.

Verify​

Ask your agent:

Search RushDB for SESSION records. Show the most recent one and summarize any linked decisions or tasks.

If you have not bootstrapped memory yet, use the Agent Memory Quickstart first.

Next Steps​