Anthropic

Installation

Install Vitamem and both peer dependencies. Anthropic does not offer an embedding API, so the adapter uses the OpenAI SDK for embeddings.

npm install vitamem @anthropic-ai/sdk openai

Quick Setup

import { createVitamem } from "vitamem";

const mem = await createVitamem({
  provider: "anthropic",
  apiKey: process.env.ANTHROPIC_API_KEY!,
  storage: "ephemeral",
});

Note

When using the provider: "anthropic" shortcut, the same apiKey value is used for both the Anthropic chat API and the OpenAI embedding API. If your keys differ, use the adapter factory directly (see below).

Adapter Factory

For full control, use createAnthropicAdapter:

import { createAnthropicAdapter, createVitamem } from "vitamem";

const llm = createAnthropicAdapter({
  apiKey: process.env.ANTHROPIC_API_KEY!,
  embeddingApiKey: process.env.OPENAI_API_KEY!,
  chatModel: "claude-sonnet-4-20250514",
  embeddingModel: "text-embedding-3-small",
});

const mem = await createVitamem({
  llm,
  storage: "ephemeral",
});

Options

Option	Type	Default	Description
apiKey	string	required	Your Anthropic API key.
chatModel	string	"claude-sonnet-4-20250514"	Model used for chat and memory extraction.
embeddingApiKey	string	required	OpenAI API key used for embeddings.
embeddingModel	string	"text-embedding-3-small"	OpenAI model used for embeddings.
baseUrl	string	undefined	Override the Anthropic API base URL.
extractionPrompt	string	Built-in prompt	Custom prompt for memory extraction. Must include a {conversation} placeholder.

Why Two API Keys?

Anthropic’s API provides chat completions but does not offer a text embedding endpoint. Since Vitamem needs embeddings for semantic memory search, the Anthropic adapter delegates embedding calls to OpenAI.

Chat + Extraction  -->  Anthropic API  (uses apiKey)
Embeddings         -->  OpenAI API     (uses embeddingApiKey)

This means you need active accounts with both providers. If you prefer a single-provider setup, consider the OpenAI adapter or bringing your own embeddings via a custom adapter.

System Prompt Handling

The Anthropic Messages API treats system prompts differently from OpenAI. Instead of accepting { role: "system" } messages in the messages array, Anthropic expects a top-level system parameter.

The adapter handles this automatically. When Vitamem passes messages containing system-role entries (for example, injected memory context when autoRetrieve is enabled), the adapter:

1. Separates system messages from the conversation
2. Joins all system message content into a single system parameter
3. Passes only user and assistant messages in the messages array

// This works transparently -- system prompts are handled for you
const { reply } = await mem.chat({
  threadId: thread.id,
  message: "How has my blood pressure been trending?",
  systemPrompt: "You are a supportive health companion. Be empathetic and precise.",
});

You do not need to do anything special for system prompts. The adapter handles the conversion.

Choosing Models

Chat Models

Any model available through the Anthropic Messages API works:

Model	Best For
claude-sonnet-4-20250514	Default. Strong reasoning, good extraction quality.
claude-haiku-35-20241022	Faster responses, lower cost.
claude-opus-4-20250514	Maximum capability for complex health conversations.

const llm = createAnthropicAdapter({
  apiKey: process.env.ANTHROPIC_API_KEY!,
  embeddingApiKey: process.env.OPENAI_API_KEY!,
  chatModel: "claude-opus-4-20250514",
});

Embedding Models

Since embeddings go through OpenAI, see the OpenAI provider docs for available embedding models.

Caution

Changing the embedding model after memories have been stored will produce vectors in a different space. Similarity search across old and new embeddings will not work correctly.

Custom Extraction Prompt

The Anthropic adapter includes a health-focused extraction prompt with detailed category guidance (medical conditions, medications, vital signs, lifestyle, goals, care team, and support preferences). You can override it:

const llm = createAnthropicAdapter({
  apiKey: process.env.ANTHROPIC_API_KEY!,
  embeddingApiKey: process.env.OPENAI_API_KEY!,
  extractionPrompt: `Extract key facts from this health conversation.

Conversation:
{conversation}

Return a JSON array only (no markdown, no explanation):
[{ "content": "brief factual statement", "source": "confirmed" | "inferred" }]`,
});

Using a Custom Anthropic Base URL

If you are routing through a proxy or using a self-hosted deployment, set baseUrl:

const llm = createAnthropicAdapter({
  apiKey: process.env.ANTHROPIC_API_KEY!,
  embeddingApiKey: process.env.OPENAI_API_KEY!,
  baseUrl: "https://your-proxy.example.com/v1",
});

This only affects the Anthropic chat endpoint. Embedding requests always go directly to the OpenAI API.

Streaming

The Anthropic adapter supports streaming via the Claude Messages API. Use chatStream() or chatWithUserStream() on the Vitamem instance:

const { stream } = await mem.chatStream({
  threadId: thread.id,
  message: "How has my blood pressure been trending?",
});

for await (const chunk of stream) {
  process.stdout.write(chunk);
}

System prompt handling works identically in streaming mode — system messages are separated and passed as the top-level system parameter automatically.

See Streaming Output for the full guide.

Peer Dependencies

The Anthropic adapter requires two peer dependencies:

{
  "peerDependencies": {
    "@anthropic-ai/sdk": ">=0.30.0",
    "openai": ">=4.0.0"
  }
}

Both SDKs are lazy-loaded at runtime. If either is missing, you will get a module resolution error when the adapter is first used.

Next Steps

• OpenAI Provider — single-provider setup with OpenAI for everything
• Ollama Provider — run models locally with zero config
• Custom LLM Adapter — implement the interface for any provider