getmnemo-anthropic

Mnemo helper for the Anthropic SDK. Wraps messages.create with a single memory tool (search + add), prompt caching on the system block, and optional extended thinking — the recommended setup for any long-lived Claude assistant.

Install

npm install getmnemo-anthropic @anthropic-ai/sdk

getmnemo (the core SDK) is pulled in automatically as a dependency. You only need @anthropic-ai/sdk because you construct the Anthropic client yourself and pass it in.

Get your Mnemo API key and workspace ID from app.mnemohq.com/settings/api-keys.

Quickstart (30 seconds)

import Anthropic from "@anthropic-ai/sdk";
import { Mnemo } from "getmnemo";
import { withMemoryTool } from "getmnemo-anthropic";

const claude = new Anthropic();
const memory = new Mnemo({
  apiKey: process.env.GETMNEMO_API_KEY!,
  workspaceId: process.env.GETMNEMO_WORKSPACE_ID!,
});

const agent = withMemoryTool({
  client: claude,
  getmnemo: memory,
  model: "claude-sonnet-4-6",
  thinkingBudgetTokens: 4000,
});

const result = await agent.run({
  system: "You are a personal assistant. Use memory before answering.",
  messages: [{ role: "user", content: "What did I tell you about my schedule?" }],
  metadata: { userId: "user-42" },
});

console.log(result.text);
console.log("memory tool calls:", result.memoryToolCalls);

What the wrapper does

• Exposes one tool memory with action: "search" | "add" — the model decides when to recall and when to remember.
• Runs the tool-use loop for you (default cap: 6 iterations).
• Adds cache_control: { type: "ephemeral" } to the system prompt so long stable instructions hit the prompt cache.
• Optionally enables extended thinking with a configurable token budget.
• Returns the final transcript so you can persist it elsewhere.

API

withMemoryTool(options) returns a wrapper with a single run(input) method.

withMemoryTool(options)

| Option | Type | Default | Description | | --- | --- | --- | --- | | client | Anthropic | — | An @anthropic-ai/sdk client (anything with messages.create). | | getmnemo | Mnemo | — | A Mnemo instance from getmnemo. | | model | string | "claude-sonnet-4-6" | Default model; override per call. | | maxTokens | number | 1024 | Default max_tokens; override per call. | | searchLimit | number | 5 | Default top-k for memory search (clamped to 1–50). | | cacheSystem | boolean | true | Wrap the system prompt with cache_control: { type: "ephemeral" }. | | thinkingBudgetTokens | number | — | Enable extended thinking with this token budget. | | metadata | Record<string, unknown> | {} | Static metadata merged into every persisted memory. | | maxIterations | number | 6 | Hard cap on tool-use loop iterations. |

run(input)

| Field | Type | Description | | --- | --- | --- | | messages | ChatMessage[] | Conversation so far (required). | | system | string | Optional system prompt. | | model | string | Per-call model override. | | maxTokens | number | Per-call max_tokens override. | | metadata | Record<string, unknown> | Per-call metadata merged into persisted memories. |

Returns:

| Field | Type | Description | | --- | --- | --- | | text | string | Final assistant text, concatenated across blocks. | | responses | MessagesCreateResult[] | Every raw API response from the tool-use loop. | | memoryToolCalls | number | How many times the memory tool ran. | | messages | ChatMessage[] | Final messages array after the loop — store this as your transcript. |

The package also exports MEMORY_TOOL_NAME (the literal tool name, "memory") and the WithMemoryTool* / WithMemoryRun* types.

Links

• Docs: mnemohq.com
• API keys: app.mnemohq.com/settings/api-keys

License

MIT