Vercel AI SDK integration

@agentmem/vercel-ai-provider plugs AgentMem into Vercel’s AI SDK . Three patterns — pick one or stack them.


npm install @agentmem/vercel-ai-provider ai

ai and zod are peer deps; if you already use the AI SDK you have them.

Pattern A — middleware (auto-recall)

Wrap any @ai-sdk/* model. Memories auto-inject on every call.


import { generateText, wrapLanguageModel } from 'ai'
import { anthropic } from '@ai-sdk/anthropic'
import { createAgentMemMiddleware } from '@agentmem/vercel-ai-provider'
 
const cfg = { apiKey: process.env.AGENTMEM_API_KEY!, agentId: 'support-bot' }
 
const model = wrapLanguageModel({
  model:      anthropic('claude-opus-4-7'),
  middleware: createAgentMemMiddleware(cfg),
})
 
const { text } = await generateText({ model, prompt: 'How do we handle escalations?' })

If memory search fails (network, auth), the middleware logs to stderr and lets the call proceed without injection — the model never blocks on AgentMem.

Pattern B — tools

Let the agent decide when to call memorize / recall.


import { generateText } from 'ai'
import { anthropic } from '@ai-sdk/anthropic'
import { agentmemMemorize, agentmemRecall } from '@agentmem/vercel-ai-provider'
 
await generateText({
  model:  anthropic('claude-opus-4-7'),
  tools:  {
    memorize: agentmemMemorize(cfg),
    recall:   agentmemRecall(cfg),
  },
  prompt: '...',
})

Pattern C — helpers (manual control)


import {
  addMemories,
  retrieveMemories,
  searchMemories,
} from '@agentmem/vercel-ai-provider'
 
await addMemories('User prefers email.', cfg)
 
const memories = await retrieveMemories(userMessage, cfg)
const system   = `${memories}\n\nYou are a helpful assistant.`
 
const hits = await searchMemories('what was decided?', cfg)

These are name-compatible with @mem0/vercel-ai-provider — switching is a one-line import change plus an env-var rename.

Drop-in migration from `@mem0/vercel-ai-provider`


- import { addMemories, retrieveMemories } from '@mem0/vercel-ai-provider'
+ import { addMemories, retrieveMemories } from '@agentmem/vercel-ai-provider'
 
- const cfg = { mem0ApiKey: process.env.MEM0_API_KEY, user_id: 'u1' }
+ const cfg = { apiKey: process.env.AGENTMEM_API_KEY!, agentId: 'u1' }

We don’t wrap every model SDK — we apply at the AI SDK level so any @ai-sdk/* provider works.

Config

Field	Default	Notes
`apiKey`	—	Required.
`agentId`	—	Required on the config or per-call.
`baseUrl`	hosted	Self-host override.
`scope`	private	Default scope for writes / searches.
`workflowId`	—	Default workflow.
`topK`	5	Hits per recall.
`rerank`	false	Re-rank with Gemini (+500-1500ms).
`minScore`	0	Drop hits below this score.
`prefix`	default	Custom text before injected memories.