@maximem/synap-claude-agent

Synap memory integration for Anthropic's Claude Agent SDK (TypeScript).

Python sibling published as synap-claude-agent.

Install

npm install @maximem/synap-claude-agent @anthropic-ai/claude-agent-sdk zod

Two plug points

1. Hooks — automatic context injection

import { query } from "@anthropic-ai/claude-agent-sdk";
import { createSynapHooks } from "@maximem/synap-claude-agent";

const sdk = /* your Synap SDK instance */;

for await (const message of query({
  prompt: "What did I tell you about my trial?",
  options: {
    hooks: createSynapHooks({ sdk, userId: "alice", customerId: "acme" }),
  },
})) {
  console.log(message);
}

2. MCP tools — explicit read/write

import { createSynapMcpServer } from "@maximem/synap-claude-agent";

const options = {
  mcpServers: { synap: createSynapMcpServer({ sdk, userId: "alice" }) },
  allowedTools: ["mcp__synap__synap_search", "mcp__synap__synap_remember"],
};

Use both together for automatic context injection plus explicit agent read/write.

Error policy

• Hooks never throw — SDK failures log and fall through (no context injected, no prompt recorded).
• synap_search returns a "no context available" message on SDK failure — keeps the agent loop alive.
• synap_remember returns isError: true on ingestion failure — silent drops would hide outages, so failures surface to the agent explicitly.