Neurophic
SDK

Vercel AI SDK tools

Easily integrate memory into your models with the Vercel AI SDK

@neurophic/ai-sdk wraps the JavaScript SDK for the Vercel AI SDK.

Install

npm install neurophic @neurophic/ai-sdk ai zod

createTools

import { Neurophic } from 'neurophic';
import { createTools } from '@neurophic/ai-sdk';

const client = new Neurophic({ apiKey: process.env.NEUROPHIC_API_KEY });
const { remember, recall } = createTools({
  client,
  identifier: 'user-123',
  name: 'Ada',
});

createTools returns two AI SDK tools you can pass straight to streamText or generateText:

Each tool carries a description and a zod input schema, so the model knows when and how to call it; invalid input is rejected before any request is made.

Example

import { Neurophic } from 'neurophic';
import { createTools } from '@neurophic/ai-sdk';
import { streamText, stepCountIs } from 'ai';
import { openai } from '@ai-sdk/openai';

const client = new Neurophic({ apiKey: process.env.NEUROPHIC_API_KEY });
const { remember, recall } = createTools({ client, identifier: 'user-123' });

const system = 'You are a travel assistant. Use recall before answering and remember what the user shares.';

const result = streamText({
  model: openai('gpt-4o'),
  system,
  tools: { remember, recall },
  stopWhen: stepCountIs(3),
  messages: [
    { role: 'user', content: 'Remember that I prefer window seats on long flights.' },
  ],
});

Learn more under API reference > Ingest and API reference > Retrieve.

Warning

If a call fails, the tool's execute throws a NeurophicError and the AI SDK reports it as a tool error for that step.

withContext

withContext calls client.context and returns a system prompt string with the context appended. Use it when you want memory in front of the model on every turn without giving it a tool.

import { Neurophic } from 'neurophic';
import { withContext } from '@neurophic/ai-sdk';
import { streamText } from 'ai';
import { openai } from '@ai-sdk/openai';

const client = new Neurophic({ apiKey: process.env.NEUROPHIC_API_KEY });
const userMessage = 'Can you find me a flight to Lisbon next month?';

const system = await withContext(client, {
  identifier: 'user-123',
  query: userMessage,
  system: 'You are a travel assistant.',
});

const result = streamText({
  model: openai('gpt-4o'),
  system,
  messages: [{ role: 'user', content: userMessage }],
});

withContext takes:

  • query. The text to fetch context for, usually the latest user message.
  • system. Optional; your own system prompt.
  • template. Optional (context, system) => string; the default appends a ## Context from memory heading and the context lines to system, or returns just that block when system is not set.

With the default template and the example above, the returned string is:

You are a travel assistant.

## Context from memory
- Prefers window seats on long flights. (memory)

Warning

withContext returns system unchanged (or an empty string when system is not set) in two cases: when nothing relevant is found, and when the request fails for any reason. Errors are not thrown.

Learn more under API reference > Context.

JavaScript

Learn about how to install the SDK, create a client and call every endpoint

Changelog

What changed in the API, the SDK and these docs