Quickstart

Build a chatbot that remembers users across conversations. We’ll use the Vercel AI SDK with Next.js, but Recall works with any framework.

1. Install Packages

npm install @youcraft/recall @youcraft/recall-ai-sdk \
  @youcraft/recall-adapter-sqlite @youcraft/recall-embeddings-openai \
  @youcraft/recall-extractor-openai ai @ai-sdk/openai

2. Create Memory Instance

lib/memory.ts

import { createMemory } from '@youcraft/recall'
import { createRecall } from '@youcraft/recall-ai-sdk'
import { sqliteAdapter } from '@youcraft/recall-adapter-sqlite'
import { openaiEmbeddings } from '@youcraft/recall-embeddings-openai'
import { openaiExtractor } from '@youcraft/recall-extractor-openai'

export const memory = createMemory({
  db: sqliteAdapter({ filename: 'memories.db' }),
  embeddings: openaiEmbeddings({ apiKey: process.env.OPENAI_API_KEY! }),
  extractor: openaiExtractor({ apiKey: process.env.OPENAI_API_KEY! }),
})

export const recall = createRecall({ memory })

3. Create Chat Endpoint

app/api/chat/route.ts

import { openai } from '@ai-sdk/openai'
import { streamText, convertToModelMessages } from 'ai'
import { recall } from '@/lib/memory'

export async function POST(req: Request) {
  const { messages } = await req.json()

  const result = streamText({
    model: recall(openai('gpt-5-nano'), { userId: 'user_123' }),
    system: 'You are a helpful assistant.',
    messages: convertToModelMessages(messages),
  })

  return result.toDataStreamResponse()
}

That’s it. Your chatbot now:

• Queries relevant memories before each response
• Injects them into the system prompt automatically

4. Add Memory Extraction

Memory extraction can be slow since it involves an LLM call. To avoid delaying responses, offload extraction to a background job using something like Inngest:

lib/memory.ts

export const recall = createRecall({
  memory,
  onExtract: async ({ messages, userId }) => {
    await inngest.send({
      name: 'memory/extract',
      data: { messages, userId },
    })
  },
})

Tip

For quick testing, you can call memory.extract(conversation, { userId }) directly in onExtract instead of using a background job.

What Gets Extracted?

When a user says:

“I’m a software engineer at Acme Corp. I mostly work with TypeScript on backend services.”

Recall extracts:

• User is a software engineer
• User works at Acme Corp
• User works with TypeScript
• User focuses on backend services

These are stored as separate, queryable memories.

What Gets Injected?

When the user asks something, relevant memories are injected into the prompt:

<memories>
- User is a software engineer at Acme Corp
- User works with TypeScript
- User focuses on backend services
</memories>

You are a helpful assistant.

The AI now has context without you managing it manually.

Memory Consolidation

Recall is smart about updates. If you already have “User’s name is John” and extract “User’s full name is John Doe”, it updates the existing memory instead of creating a duplicate.

// First conversation
await memory.extract('User: My name is John', { userId })
// Stores: "User's name is John"

// Later conversation
await memory.extract('User: My full name is John Doe', { userId })
// Updates to: "User's name is John Doe"

Core API

Need more control? Use the core API directly:

import { memory } from '@/lib/memory'

// Extract memories from text
await memory.extract(conversation, { userId: 'user_123' })

// Query with natural language
const relevant = await memory.query('What does the user do?', {
  userId: 'user_123',
  limit: 5,
})

// CRUD operations
const all = await memory.list('user_123')
const one = await memory.get('memory_id')
await memory.update('memory_id', { content: 'Updated' })
await memory.delete('memory_id')
await memory.clear('user_123')

Next Steps

• AI SDK Integration — Full API reference and advanced options
• Examples — Complete working examples on GitHub