Claude Agent SDK integration

@agentmem/claude-agent is the integration for Anthropic’s Claude Agent SDK . It gives you two ways to wire AgentMem into your agent — use either or both.


npm install @agentmem/claude-agent @anthropic-ai/claude-agent-sdk zod

Two integration patterns

Pattern A — in-process MCP server (agent calls tools when it decides)


import { query } from '@anthropic-ai/claude-agent-sdk'
import { agentmemMcpServer } from '@agentmem/claude-agent'
 
const cfg = { apiKey: process.env.AGENTMEM_API_KEY!, agentId: 'bot' }
 
const result = query({
  prompt: '...',
  options: {
    mcpServers: { agentmem: agentmemMcpServer(cfg) },
  },
})

Agent gets 8 memory tools (memory_write, memory_search, memory_get, memory_delete, memory_check_conflicts, scratch_set, scratch_get, workflow_summarize). No separate process — the server runs in your Node.

Pattern B — recall hook (memories auto-inject every turn)


import { query } from '@anthropic-ai/claude-agent-sdk'
import { agentmemRecallHook } from '@agentmem/claude-agent'
 
const cfg = { apiKey: process.env.AGENTMEM_API_KEY!, agentId: 'bot' }
 
query({
  prompt: '...',
  options: {
    hooks: {
      UserPromptSubmit: [{ hooks: [agentmemRecallHook(cfg)] }],
    },
  },
})

Every user prompt triggers a memory search; the top hits are spliced into the agent’s context as additionalContext. The agent never has to call a tool — memories just appear. If search fails (network, auth), the hook logs to stderr and returns empty so the turn proceeds.

Recommended — use both

Hook gives passive context every turn; tools let the agent dig deeper or write durable facts.


query({
  prompt: '...',
  options: {
    mcpServers: { agentmem: agentmemMcpServer(cfg) },
    hooks:      { UserPromptSubmit: [{ hooks: [agentmemRecallHook(cfg)] }] },
  },
})

Recall hook options

Option	Default	Purpose
`apiKey`	—	Required.
`agentId`	—	Required.
`baseUrl`	hosted	Self-host override.
`workflowId`	—	Limit to a workflow.
`scope`	all	Limit to `private`/`team`/`global`.
`topK`	5	Number of memories per turn.
`rerank`	false	Re-score with Gemini (+500-1500ms).
`minScore`	0	Skip hits below this score.
`prefix`	default preamble	Custom text before the memory block.

Comparison to `@agentmem/mcp`

@agentmem/mcp is a standalone stdio server for client apps that spawn MCP processes (Claude Desktop, Cursor).
@agentmem/claude-agent runs in your Node code that calls query() directly, no separate process — plus the recall hook.

Use both if you’re building both a desktop client and a programmatic agent.